Note

Go to the end to download the full example code.

Export with dynamic dimensions in {0,1}

The first version of torch.export.export() did not support any tensor with a dimension equal to 0, 1 if the dimension was expected to be dynamic. The latest versions offers more options. Let’s check it works. The experiments consists in exporting the model with different sets of inputs and checking the exported models works with all set of inputs.

Available input sets

import itertools
from tqdm import tqdm
import numpy as np
import pandas
import torch
from onnx_diagnostic import doc
from onnx_diagnostic.helpers import max_diff, string_type
from onnx_diagnostic.helpers.torch_helper import torch_deepcopy
from onnx_diagnostic.torch_models.hghub.model_inputs import get_untrained_model_with_inputs
from onnx_diagnostic.torch_export_patches.patch_inputs import use_dyn_not_str
from onnx_diagnostic.torch_export_patches import (
    torch_export_patches,
    register_additional_serialization_functions,
)


data = get_untrained_model_with_inputs("arnir0/Tiny-LLM", add_second_input=True)
model, dynamic_shapes = data["model"], data["dynamic_shapes"]

The trained model can be obtained with:

MODEL_NAME = "arnir0/Tiny-LLM"
tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)
model = transformers.AutoModelForCausalLM.from_pretrained(MODEL_NAME)

input_sets = {k: v for k, v in data.items() if k.startswith("inputs")}

for k, v in input_sets.items():
    print(f"{k:20}: {string_type(v, with_shape=True)}")

inputs              : dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#1[T1s2x1x30x96], value_cache=#1[T1s2x1x30x96]))
inputs2             : dict(input_ids:T7s3x4,attention_mask:T7s3x35,position_ids:T7s3x4,past_key_values:DynamicCache(key_cache=#1[T1s3x1x31x96], value_cache=#1[T1s3x1x31x96]))
inputs_empty_cache  : dict(input_ids:T7s2x3,attention_mask:T7s2x3,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#1[T1s2x1x0x96], value_cache=#1[T1s2x1x0x96]))
inputs_batch1       : dict(input_ids:T7s1x3,attention_mask:T7s1x33,position_ids:T7s1x3,past_key_values:DynamicCache(key_cache=#1[T1s1x1x30x96], value_cache=#1[T1s1x1x30x96]))

The dynamic shapes are:

print(f"dynamic_shapes: {string_type(dynamic_shapes)}")

dynamic_shapes: dict(input_ids:{0:DYN(batch),1:DYN(seq_length)},attention_mask:{0:DYN(batch),1:DYN(cache+seq)},position_ids:{0:DYN(batch),1:DYN(cache+seq)},past_key_values:#2[#1[{0:DYN(batch),2:DYN(cache_length)}],#1[{0:DYN(batch),2:DYN(cache_length)}]])

dynamic_shapes = use_dyn_not_str(dynamic_shapes)
print(f"dynamic_shapes: {string_type(dynamic_shapes)}")

dynamic_shapes: dict(input_ids:{0:DYNAMIC,1:DYNAMIC},attention_mask:{0:DYNAMIC,1:DYNAMIC},position_ids:{0:DYNAMIC,1:DYNAMIC},past_key_values:#2[#1[{0:DYNAMIC,2:DYNAMIC}],#1[{0:DYNAMIC,2:DYNAMIC}]])

Let’s check they all work and compute the expected values. We use deepcopy because caches are usually modified inplace.

expected = {}
for k, v in input_sets.items():
    expected[k] = model(**torch_deepcopy(v))
    print(f"{k:20}: {string_type(expected[k], with_shape=True)}")

inputs              : CausalLMOutputWithPast(logits:T1s2x3x32000,past_key_values:DynamicCache(key_cache=#1[T1s2x1x33x96], value_cache=#1[T1s2x1x33x96]))
inputs2             : CausalLMOutputWithPast(logits:T1s3x4x32000,past_key_values:DynamicCache(key_cache=#1[T1s3x1x35x96], value_cache=#1[T1s3x1x35x96]))
inputs_empty_cache  : CausalLMOutputWithPast(logits:T1s2x3x32000,past_key_values:DynamicCache(key_cache=#1[T1s2x1x3x96], value_cache=#1[T1s2x1x3x96]))
inputs_batch1       : CausalLMOutputWithPast(logits:T1s1x3x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x33x96], value_cache=#1[T1s1x1x33x96]))

Export with options

We try to export with the following options:

Some function first.

def export_model(
    model,
    dynamic_shapes,
    inputs,
    cache=False,
    oblivious=False,
    rt=False,
    cache_patch=False,
    strict=False,
):
    if cache and not cache_patch:
        with register_additional_serialization_functions(patch_transformers=True):
            return export_model(
                model, dynamic_shapes, inputs, oblivious=oblivious, rt=rt, strict=strict
            )
    if cache_patch:
        with torch_export_patches(
            patch_torch=cache_patch in ("all", "torch", True, 1),
            patch_transformers=cache_patch in ("all", "transformers", True, 1),
        ):
            return export_model(
                model, dynamic_shapes, inputs, oblivious=oblivious, rt=rt, strict=strict
            )
    if oblivious:
        with torch.fx.experimental._config.patch(backed_size_oblivious=True):
            return export_model(model, dynamic_shapes, inputs, rt=rt, strict=strict)
    return torch.export.export(
        model,
        (),
        inputs,
        dynamic_shapes=dynamic_shapes,
        strict=strict,
        prefer_deferred_runtime_asserts_over_guards=rt,
    )


def try_export_model(
    model,
    dynamic_shapes,
    inputs,
    cache=False,
    oblivious=False,
    rt=False,
    cache_patch=False,
    strict=False,
):
    try:
        return export_model(
            model,
            dynamic_shapes,
            inputs,
            cache=cache,
            oblivious=oblivious,
            rt=rt,
            cache_patch=cache_patch,
            strict=strict,
        )
    except Exception as e:
        return e


def validation(ep, input_sets, expected):
    mod = ep.module()
    for k, v in input_sets.items():
        try:
            got = mod(**torch_deepcopy(v))
        except Exception as e:
            yield k, e
            continue
        yield k, max_diff(expected[k], got, verbose=0)

The main loop

results = []

possibilities = [*[[0, 1] for _ in range(5)], list(input_sets)]
possibilities[1] = [0, "all", "torch", "transformers"]
with tqdm(list(itertools.product(*possibilities))) as pbar:
    for cache, cache_patch, strict, oblivious, rt, inputs in pbar:
        if cache_patch and not cache:
            # patches include caches.
            continue
        kwargs = dict(
            cache=cache, cache_patch=cache_patch, strict=strict, oblivious=oblivious, rt=rt
        )
        legend = "-".join(
            (k if isinstance(v, int) else f"{k}:{v}") for k, v in kwargs.items() if v
        )
        legend = f"{legend}/{inputs}"
        pbar.set_description(f"{legend} EXPORT")

        # export
        ep = try_export_model(
            model, dynamic_shapes, torch_deepcopy(input_sets[inputs]), **kwargs
        )
        if isinstance(ep, Exception):
            obs = {
                **kwargs,
                "export_with": inputs,
                "EXPORT": 0,
                "ERR-EXPORT": str(ep).split("\n")[0],
            }
            results.append(obs)
            continue

        pbar.set_description(f"{legend} VALIDATE")
        common = {**kwargs, "export_with": inputs, "EXPORT": 1}
        for inp, res in validation(ep, input_sets, expected):
            if isinstance(res, Exception):
                obs = {
                    **common,
                    "run_with": inp,
                    "ERR-RUN": str(res).split("\n")[0],
                    "WORKS": 0,
                }
            else:
                obs = {
                    **common,
                    "run_with": inp,
                    "WORKS": int(~np.isnan(res["abs"]) and res["abs"] < 1e-3),
                }
            results.append(obs)

  0%|          | 0/256 [00:00<?, ?it/s]
/inputs EXPORT:   0%|          | 0/256 [00:00<?, ?it/s]
/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
rt/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
rt/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
rt/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
rt/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious-rt/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious-rt/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious-rt/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
oblivious-rt/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-rt/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-rt/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-rt/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-rt/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious-rt/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious-rt/inputs2 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious-rt/inputs_empty_cache EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
strict-oblivious-rt/inputs_batch1 EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
cache/inputs EXPORT:   0%|          | 1/256 [00:00<01:05,  3.90it/s]
cache/inputs EXPORT:  50%|█████     | 129/256 [00:00<00:00, 161.61it/s]
cache/inputs2 EXPORT:  50%|█████     | 129/256 [00:00<00:00, 161.61it/s]
cache/inputs_empty_cache EXPORT:  50%|█████     | 129/256 [00:01<00:00, 161.61it/s]
cache/inputs_batch1 EXPORT:  50%|█████     | 129/256 [00:01<00:00, 161.61it/s]
cache-rt/inputs EXPORT:  50%|█████     | 129/256 [00:01<00:00, 161.61it/s]
cache-rt/inputs2 EXPORT:  50%|█████     | 129/256 [00:01<00:00, 161.61it/s]
cache-rt/inputs_empty_cache EXPORT:  50%|█████     | 129/256 [00:01<00:00, 161.61it/s]
cache-rt/inputs_batch1 EXPORT:  50%|█████     | 129/256 [00:02<00:00, 161.61it/s]
cache-oblivious/inputs EXPORT:  50%|█████     | 129/256 [00:02<00:00, 161.61it/s]
cache-oblivious/inputs2 EXPORT:  50%|█████     | 129/256 [00:02<00:00, 161.61it/s]
cache-oblivious/inputs_empty_cache EXPORT:  50%|█████     | 129/256 [00:02<00:00, 161.61it/s]
cache-oblivious/inputs_batch1 EXPORT:  50%|█████     | 129/256 [00:02<00:00, 161.61it/s]
cache-oblivious-rt/inputs EXPORT:  50%|█████     | 129/256 [00:03<00:00, 161.61it/s]
cache-oblivious-rt/inputs2 EXPORT:  50%|█████     | 129/256 [00:03<00:00, 161.61it/s]
cache-oblivious-rt/inputs2 EXPORT:  55%|█████▌    | 142/256 [00:03<00:03, 33.24it/s]
cache-oblivious-rt/inputs_empty_cache EXPORT:  55%|█████▌    | 142/256 [00:03<00:03, 33.24it/s]
cache-oblivious-rt/inputs_batch1 EXPORT:  55%|█████▌    | 142/256 [00:03<00:03, 33.24it/s]
cache-strict/inputs EXPORT:  55%|█████▌    | 142/256 [00:03<00:03, 33.24it/s]
cache-strict/inputs2 EXPORT:  55%|█████▌    | 142/256 [00:04<00:03, 33.24it/s]
cache-strict/inputs_empty_cache EXPORT:  55%|█████▌    | 142/256 [00:04<00:03, 33.24it/s]
cache-strict/inputs_batch1 EXPORT:  55%|█████▌    | 142/256 [00:05<00:03, 33.24it/s]
cache-strict-rt/inputs EXPORT:  55%|█████▌    | 142/256 [00:05<00:03, 33.24it/s]
cache-strict-rt/inputs EXPORT:  58%|█████▊    | 149/256 [00:06<00:06, 16.22it/s]
cache-strict-rt/inputs2 EXPORT:  58%|█████▊    | 149/256 [00:06<00:06, 16.22it/s]
cache-strict-rt/inputs_empty_cache EXPORT:  58%|█████▊    | 149/256 [00:06<00:06, 16.22it/s]
cache-strict-rt/inputs_batch1 EXPORT:  58%|█████▊    | 149/256 [00:06<00:06, 16.22it/s]
cache-strict-oblivious/inputs EXPORT:  58%|█████▊    | 149/256 [00:07<00:06, 16.22it/s]
cache-strict-oblivious/inputs2 EXPORT:  58%|█████▊    | 149/256 [00:07<00:06, 16.22it/s]
cache-strict-oblivious/inputs2 EXPORT:  60%|██████    | 154/256 [00:07<00:08, 11.78it/s]
cache-strict-oblivious/inputs_empty_cache EXPORT:  60%|██████    | 154/256 [00:07<00:08, 11.78it/s]
cache-strict-oblivious/inputs_batch1 EXPORT:  60%|██████    | 154/256 [00:08<00:08, 11.78it/s]
cache-strict-oblivious-rt/inputs EXPORT:  60%|██████    | 154/256 [00:08<00:08, 11.78it/s]
cache-strict-oblivious-rt/inputs EXPORT:  61%|██████▏   | 157/256 [00:08<00:10,  9.77it/s]
cache-strict-oblivious-rt/inputs2 EXPORT:  61%|██████▏   | 157/256 [00:08<00:10,  9.77it/s]
cache-strict-oblivious-rt/inputs_empty_cache EXPORT:  61%|██████▏   | 157/256 [00:09<00:10,  9.77it/s]
cache-strict-oblivious-rt/inputs_empty_cache EXPORT:  62%|██████▏   | 159/256 [00:09<00:11,  8.51it/s]
cache-strict-oblivious-rt/inputs_batch1 EXPORT:  62%|██████▏   | 159/256 [00:09<00:11,  8.51it/s]
cache-cache_patch:all/inputs EXPORT:  62%|██████▏   | 159/256 [00:09<00:11,  8.51it/s]
cache-cache_patch:all/inputs VALIDATE:  62%|██████▏   | 159/256 [00:09<00:11,  8.51it/s]
cache-cache_patch:all/inputs2 EXPORT:  62%|██████▏   | 159/256 [00:09<00:11,  8.51it/s]
cache-cache_patch:all/inputs2 VALIDATE:  62%|██████▏   | 159/256 [00:10<00:11,  8.51it/s]
cache-cache_patch:all/inputs2 VALIDATE:  63%|██████▎   | 162/256 [00:11<00:15,  6.23it/s]
cache-cache_patch:all/inputs_empty_cache EXPORT:  63%|██████▎   | 162/256 [00:11<00:15,  6.23it/s]
cache-cache_patch:all/inputs_empty_cache VALIDATE:  63%|██████▎   | 162/256 [00:11<00:15,  6.23it/s]
cache-cache_patch:all/inputs_batch1 EXPORT:  63%|██████▎   | 162/256 [00:11<00:15,  6.23it/s]
cache-cache_patch:all/inputs_batch1 VALIDATE:  63%|██████▎   | 162/256 [00:12<00:15,  6.23it/s]
cache-cache_patch:all/inputs_batch1 VALIDATE:  64%|██████▍   | 164/256 [00:12<00:20,  4.44it/s]
cache-cache_patch:all-rt/inputs EXPORT:  64%|██████▍   | 164/256 [00:12<00:20,  4.44it/s]
cache-cache_patch:all-rt/inputs VALIDATE:  64%|██████▍   | 164/256 [00:13<00:20,  4.44it/s]
cache-cache_patch:all-rt/inputs VALIDATE:  64%|██████▍   | 165/256 [00:13<00:24,  3.70it/s]
cache-cache_patch:all-rt/inputs2 EXPORT:  64%|██████▍   | 165/256 [00:13<00:24,  3.70it/s]
cache-cache_patch:all-rt/inputs2 VALIDATE:  64%|██████▍   | 165/256 [00:14<00:24,  3.70it/s]
cache-cache_patch:all-rt/inputs2 VALIDATE:  65%|██████▍   | 166/256 [00:14<00:29,  3.05it/s]
cache-cache_patch:all-rt/inputs_empty_cache EXPORT:  65%|██████▍   | 166/256 [00:14<00:29,  3.05it/s]
cache-cache_patch:all-rt/inputs_empty_cache VALIDATE:  65%|██████▍   | 166/256 [00:15<00:29,  3.05it/s]
cache-cache_patch:all-rt/inputs_empty_cache VALIDATE:  65%|██████▌   | 167/256 [00:15<00:33,  2.62it/s]
cache-cache_patch:all-rt/inputs_batch1 EXPORT:  65%|██████▌   | 167/256 [00:15<00:33,  2.62it/s]
cache-cache_patch:all-rt/inputs_batch1 VALIDATE:  65%|██████▌   | 167/256 [00:16<00:33,  2.62it/s]
cache-cache_patch:all-rt/inputs_batch1 VALIDATE:  66%|██████▌   | 168/256 [00:16<00:37,  2.34it/s]
cache-cache_patch:all-oblivious/inputs EXPORT:  66%|██████▌   | 168/256 [00:16<00:37,  2.34it/s]
cache-cache_patch:all-oblivious/inputs VALIDATE:  66%|██████▌   | 168/256 [00:17<00:37,  2.34it/s]
cache-cache_patch:all-oblivious/inputs VALIDATE:  66%|██████▌   | 169/256 [00:17<00:44,  1.95it/s]
cache-cache_patch:all-oblivious/inputs2 EXPORT:  66%|██████▌   | 169/256 [00:17<00:44,  1.95it/s]
cache-cache_patch:all-oblivious/inputs2 VALIDATE:  66%|██████▌   | 169/256 [00:17<00:44,  1.95it/s]
cache-cache_patch:all-oblivious/inputs2 VALIDATE:  66%|██████▋   | 170/256 [00:17<00:49,  1.73it/s]
cache-cache_patch:all-oblivious/inputs_empty_cache EXPORT:  66%|██████▋   | 170/256 [00:17<00:49,  1.73it/s]
cache-cache_patch:all-oblivious/inputs_empty_cache VALIDATE:  66%|██████▋   | 170/256 [00:19<00:49,  1.73it/s]
cache-cache_patch:all-oblivious/inputs_empty_cache VALIDATE:  67%|██████▋   | 171/256 [00:19<01:02,  1.36it/s]
cache-cache_patch:all-oblivious/inputs_batch1 EXPORT:  67%|██████▋   | 171/256 [00:19<01:02,  1.36it/s]
cache-cache_patch:all-oblivious/inputs_batch1 VALIDATE:  67%|██████▋   | 171/256 [00:20<01:02,  1.36it/s]
cache-cache_patch:all-oblivious/inputs_batch1 VALIDATE:  67%|██████▋   | 172/256 [00:20<01:04,  1.30it/s]
cache-cache_patch:all-oblivious-rt/inputs EXPORT:  67%|██████▋   | 172/256 [00:20<01:04,  1.30it/s]
cache-cache_patch:all-oblivious-rt/inputs VALIDATE:  67%|██████▋   | 172/256 [00:21<01:04,  1.30it/s]
cache-cache_patch:all-oblivious-rt/inputs VALIDATE:  68%|██████▊   | 173/256 [00:21<01:07,  1.22it/s]
cache-cache_patch:all-oblivious-rt/inputs2 EXPORT:  68%|██████▊   | 173/256 [00:21<01:07,  1.22it/s]
cache-cache_patch:all-oblivious-rt/inputs2 VALIDATE:  68%|██████▊   | 173/256 [00:22<01:07,  1.22it/s]
cache-cache_patch:all-oblivious-rt/inputs2 VALIDATE:  68%|██████▊   | 174/256 [00:22<01:10,  1.16it/s]
cache-cache_patch:all-oblivious-rt/inputs_empty_cache EXPORT:  68%|██████▊   | 174/256 [00:22<01:10,  1.16it/s]
cache-cache_patch:all-oblivious-rt/inputs_empty_cache VALIDATE:  68%|██████▊   | 174/256 [00:22<01:10,  1.16it/s]
cache-cache_patch:all-oblivious-rt/inputs_empty_cache VALIDATE:  68%|██████▊   | 175/256 [00:23<01:11,  1.14it/s]
cache-cache_patch:all-oblivious-rt/inputs_batch1 EXPORT:  68%|██████▊   | 175/256 [00:23<01:11,  1.14it/s]
cache-cache_patch:all-oblivious-rt/inputs_batch1 VALIDATE:  68%|██████▊   | 175/256 [00:23<01:11,  1.14it/s]
cache-cache_patch:all-oblivious-rt/inputs_batch1 VALIDATE:  69%|██████▉   | 176/256 [00:23<01:09,  1.16it/s]
cache-cache_patch:all-strict/inputs EXPORT:  69%|██████▉   | 176/256 [00:23<01:09,  1.16it/s]               ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict/inputs EXPORT:  69%|██████▉   | 177/256 [00:25<01:15,  1.04it/s]
cache-cache_patch:all-strict/inputs2 EXPORT:  69%|██████▉   | 177/256 [00:25<01:15,  1.04it/s]
cache-cache_patch:all-strict/inputs2 EXPORT:  70%|██████▉   | 178/256 [00:26<01:17,  1.01it/s]
cache-cache_patch:all-strict/inputs_empty_cache EXPORT:  70%|██████▉   | 178/256 [00:26<01:17,  1.01it/s]
cache-cache_patch:all-strict/inputs_empty_cache EXPORT:  70%|██████▉   | 179/256 [00:27<01:30,  1.17s/it]
cache-cache_patch:all-strict/inputs_batch1 EXPORT:  70%|██████▉   | 179/256 [00:27<01:30,  1.17s/it]
cache-cache_patch:all-strict/inputs_batch1 EXPORT:  70%|███████   | 180/256 [00:33<03:07,  2.46s/it]
cache-cache_patch:all-strict-rt/inputs EXPORT:  70%|███████   | 180/256 [00:33<03:07,  2.46s/it]    ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-rt/inputs EXPORT:  71%|███████   | 181/256 [00:35<02:49,  2.27s/it]
cache-cache_patch:all-strict-rt/inputs2 EXPORT:  71%|███████   | 181/256 [00:35<02:49,  2.27s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-rt/inputs2 EXPORT:  71%|███████   | 182/256 [00:36<02:38,  2.14s/it]
cache-cache_patch:all-strict-rt/inputs_empty_cache EXPORT:  71%|███████   | 182/256 [00:36<02:38,  2.14s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-rt/inputs_empty_cache EXPORT:  71%|███████▏  | 183/256 [00:38<02:29,  2.04s/it]
cache-cache_patch:all-strict-rt/inputs_batch1 EXPORT:  71%|███████▏  | 183/256 [00:38<02:29,  2.04s/it]     ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-rt/inputs_batch1 EXPORT:  72%|███████▏  | 184/256 [00:40<02:18,  1.92s/it]
cache-cache_patch:all-strict-oblivious/inputs EXPORT:  72%|███████▏  | 184/256 [00:40<02:18,  1.92s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-oblivious/inputs2 EXPORT:  72%|███████▏  | 184/256 [00:40<02:18,  1.92s/it]
cache-cache_patch:all-strict-oblivious/inputs2 EXPORT:  73%|███████▎  | 186/256 [00:42<01:39,  1.42s/it]
cache-cache_patch:all-strict-oblivious/inputs_empty_cache EXPORT:  73%|███████▎  | 186/256 [00:42<01:39,  1.42s/it]
cache-cache_patch:all-strict-oblivious/inputs_empty_cache EXPORT:  73%|███████▎  | 187/256 [00:43<01:33,  1.35s/it]
cache-cache_patch:all-strict-oblivious/inputs_batch1 EXPORT:  73%|███████▎  | 187/256 [00:43<01:33,  1.35s/it]
cache-cache_patch:all-strict-oblivious/inputs_batch1 EXPORT:  73%|███████▎  | 188/256 [00:45<01:45,  1.55s/it]
cache-cache_patch:all-strict-oblivious-rt/inputs EXPORT:  73%|███████▎  | 188/256 [00:45<01:45,  1.55s/it]    ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-oblivious-rt/inputs EXPORT:  74%|███████▍  | 189/256 [00:47<01:52,  1.68s/it]
cache-cache_patch:all-strict-oblivious-rt/inputs2 EXPORT:  74%|███████▍  | 189/256 [00:47<01:52,  1.68s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-oblivious-rt/inputs2 EXPORT:  74%|███████▍  | 190/256 [00:49<01:59,  1.80s/it]
cache-cache_patch:all-strict-oblivious-rt/inputs_empty_cache EXPORT:  74%|███████▍  | 190/256 [00:49<01:59,  1.80s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-oblivious-rt/inputs_empty_cache EXPORT:  75%|███████▍  | 191/256 [00:51<02:04,  1.91s/it]
cache-cache_patch:all-strict-oblivious-rt/inputs_batch1 EXPORT:  75%|███████▍  | 191/256 [00:51<02:04,  1.91s/it]     ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:all-strict-oblivious-rt/inputs_batch1 EXPORT:  75%|███████▌  | 192/256 [00:53<02:02,  1.91s/it]
cache-cache_patch:torch/inputs EXPORT:  75%|███████▌  | 192/256 [00:53<02:02,  1.91s/it]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch/inputs EXPORT:  75%|███████▌  | 193/256 [00:53<01:30,  1.43s/it]
cache-cache_patch:torch/inputs2 EXPORT:  75%|███████▌  | 193/256 [00:53<01:30,  1.43s/it]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch/inputs2 EXPORT:  76%|███████▌  | 194/256 [00:54<01:06,  1.08s/it]
cache-cache_patch:torch/inputs_empty_cache EXPORT:  76%|███████▌  | 194/256 [00:54<01:06,  1.08s/it]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, 0, 96]", arg17_1: "f32[s4, 1, 0, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, 0, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = cat = None
    cat_1: "f32[s4, 1, 0, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s70)" = 0 + sym_size_int;  sym_size_int = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(0, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s70)" = 0 + sym_size_int_1
    add_2: "Sym(s70)" = add_1 + 0
    sym_size_int_2: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(-s53 + s70)" = add_2 - sym_size_int_2;  add_2 = None
    gt: "Sym(-s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s70)" = sym_size_int_2 > add_1;  sym_size_int_2 = gt_1 = None
    arange_1: "i64[s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_3: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_3, device = device(type='cpu'), pin_memory = False);  sym_size_int_3 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_4, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_4, sym_size_int_1]);  unsqueeze_1 = sym_size_int_1 = expand_1 = None
    unsqueeze_2: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_5: "Sym(s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_4, sym_size_int_5]);  unsqueeze_2 = sym_size_int_4 = sym_size_int_5 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, 0, 96]", arg17_1: "f32[s4, 1, 0, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, 0, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = cat = None
    cat_1: "f32[s4, 1, 0, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s70)" = 0 + sym_size_int;  sym_size_int = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(0, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s70)" = 0 + sym_size_int_1
    add_2: "Sym(s70)" = add_1 + 0
    sym_size_int_2: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(-s53 + s70)" = add_2 - sym_size_int_2;  add_2 = None
    gt: "Sym(-s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s70)" = sym_size_int_2 > add_1;  sym_size_int_2 = gt_1 = None
    arange_1: "i64[s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_3: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_3, device = device(type='cpu'), pin_memory = False);  sym_size_int_3 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_4, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_4, sym_size_int_1]);  unsqueeze_1 = sym_size_int_1 = expand_1 = None
    unsqueeze_2: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_5: "Sym(s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_4, sym_size_int_5]);  unsqueeze_2 = sym_size_int_4 = sym_size_int_5 = expand_2 = None


cache-cache_patch:torch/inputs_empty_cache EXPORT:  76%|███████▌  | 195/256 [00:55<01:09,  1.13s/it]
cache-cache_patch:torch/inputs_batch1 EXPORT:  76%|███████▌  | 195/256 [00:55<01:09,  1.13s/it]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[1, s70]", arg14_1: "i64[1, s53]", arg15_1: "i64[1, s9]", arg16_1: "f32[1, 1, s31, 96]", arg17_1: "f32[1, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[1, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[1, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[1, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1);  arg13_1 = None
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[1, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    arange_2: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[1]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  arange_2 = None
    select: "i64[]" = torch.ops.aten.select.int(movedim, 0, 0);  movedim = None
    movedim_1: "i64[1]" = torch.ops.aten.movedim.int(arange_3, 0, 0);  arange_3 = None
    select_1: "i64[]" = torch.ops.aten.select.int(movedim_1, 0, 0);  movedim_1 = None
    movedim_2: "i64[s70]" = torch.ops.aten.movedim.int(arange, 0, 0);  arange = movedim_2 = None
    unsqueeze: "i64[1]" = torch.ops.aten.unsqueeze.default(select, 0);  select = None
    expand: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_2]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1]" = torch.ops.aten.unsqueeze.default(select_1, 0);  select_1 = None
    expand_1: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_2]);  unsqueeze_1 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_4: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s70, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_2, sym_size_int_4]);  unsqueeze_2 = sym_size_int_2 = sym_size_int_4 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[1, s70]", arg14_1: "i64[1, s53]", arg15_1: "i64[1, s9]", arg16_1: "f32[1, 1, s31, 96]", arg17_1: "f32[1, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[1, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[1, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[1, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1);  arg13_1 = None
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[1, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    arange_2: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[1]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  arange_2 = None
    select: "i64[]" = torch.ops.aten.select.int(movedim, 0, 0);  movedim = None
    movedim_1: "i64[1]" = torch.ops.aten.movedim.int(arange_3, 0, 0);  arange_3 = None
    select_1: "i64[]" = torch.ops.aten.select.int(movedim_1, 0, 0);  movedim_1 = None
    movedim_2: "i64[s70]" = torch.ops.aten.movedim.int(arange, 0, 0);  arange = movedim_2 = None
    unsqueeze: "i64[1]" = torch.ops.aten.unsqueeze.default(select, 0);  select = None
    expand: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_2]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1]" = torch.ops.aten.unsqueeze.default(select_1, 0);  select_1 = None
    expand_1: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_2]);  unsqueeze_1 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_4: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s70, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_2, sym_size_int_4]);  unsqueeze_2 = sym_size_int_2 = sym_size_int_4 = expand_2 = None


cache-cache_patch:torch/inputs_batch1 EXPORT:  77%|███████▋  | 196/256 [00:55<00:51,  1.17it/s]
cache-cache_patch:torch-rt/inputs EXPORT:  77%|███████▋  | 196/256 [00:55<00:51,  1.17it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-rt/inputs EXPORT:  77%|███████▋  | 197/256 [00:55<00:41,  1.41it/s]
cache-cache_patch:torch-rt/inputs2 EXPORT:  77%|███████▋  | 197/256 [00:55<00:41,  1.41it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-rt/inputs2 EXPORT:  77%|███████▋  | 198/256 [00:56<00:35,  1.63it/s]
cache-cache_patch:torch-rt/inputs_empty_cache EXPORT:  77%|███████▋  | 198/256 [00:56<00:35,  1.63it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, 0, 96]", arg17_1: "f32[s4, 1, 0, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, 0, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = cat = None
    cat_1: "f32[s4, 1, 0, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s70)" = 0 + sym_size_int;  sym_size_int = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(0, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s70)" = 0 + sym_size_int_1
    add_2: "Sym(s70)" = add_1 + 0
    sym_size_int_2: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(-s53 + s70)" = add_2 - sym_size_int_2;  add_2 = None
    gt: "Sym(-s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s70)" = sym_size_int_2 > add_1;  sym_size_int_2 = gt_1 = None
    arange_1: "i64[s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_3: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_3, device = device(type='cpu'), pin_memory = False);  sym_size_int_3 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_4, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_4, sym_size_int_1]);  unsqueeze_1 = sym_size_int_1 = expand_1 = None
    unsqueeze_2: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_5: "Sym(s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_4, sym_size_int_5]);  unsqueeze_2 = sym_size_int_4 = sym_size_int_5 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, 0, 96]", arg17_1: "f32[s4, 1, 0, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, 0, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = cat = None
    cat_1: "f32[s4, 1, 0, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s70)" = 0 + sym_size_int;  sym_size_int = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(0, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s70)" = 0 + sym_size_int_1
    add_2: "Sym(s70)" = add_1 + 0
    sym_size_int_2: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(-s53 + s70)" = add_2 - sym_size_int_2;  add_2 = None
    gt: "Sym(-s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s70)" = sym_size_int_2 > add_1;  sym_size_int_2 = gt_1 = None
    arange_1: "i64[s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_3: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_3, device = device(type='cpu'), pin_memory = False);  sym_size_int_3 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_4, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_4, sym_size_int_1]);  unsqueeze_1 = sym_size_int_1 = expand_1 = None
    unsqueeze_2: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_5: "Sym(s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_4, sym_size_int_5]);  unsqueeze_2 = sym_size_int_4 = sym_size_int_5 = expand_2 = None


cache-cache_patch:torch-rt/inputs_empty_cache EXPORT:  78%|███████▊  | 199/256 [00:56<00:28,  2.01it/s]
cache-cache_patch:torch-rt/inputs_batch1 EXPORT:  78%|███████▊  | 199/256 [00:56<00:28,  2.01it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[1, s70]", arg14_1: "i64[1, s53]", arg15_1: "i64[1, s9]", arg16_1: "f32[1, 1, s31, 96]", arg17_1: "f32[1, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[1, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[1, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[1, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1);  arg13_1 = None
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[1, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    arange_2: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[1]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  arange_2 = None
    select: "i64[]" = torch.ops.aten.select.int(movedim, 0, 0);  movedim = None
    movedim_1: "i64[1]" = torch.ops.aten.movedim.int(arange_3, 0, 0);  arange_3 = None
    select_1: "i64[]" = torch.ops.aten.select.int(movedim_1, 0, 0);  movedim_1 = None
    movedim_2: "i64[s70]" = torch.ops.aten.movedim.int(arange, 0, 0);  arange = movedim_2 = None
    unsqueeze: "i64[1]" = torch.ops.aten.unsqueeze.default(select, 0);  select = None
    expand: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_2]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1]" = torch.ops.aten.unsqueeze.default(select_1, 0);  select_1 = None
    expand_1: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_2]);  unsqueeze_1 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_4: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s70, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_2, sym_size_int_4]);  unsqueeze_2 = sym_size_int_2 = sym_size_int_4 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[1, s70]", arg14_1: "i64[1, s53]", arg15_1: "i64[1, s9]", arg16_1: "f32[1, 1, s31, 96]", arg17_1: "f32[1, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[1, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[1, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[1, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1);  arg13_1 = None
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[1, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    arange_2: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[1]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  arange_2 = None
    select: "i64[]" = torch.ops.aten.select.int(movedim, 0, 0);  movedim = None
    movedim_1: "i64[1]" = torch.ops.aten.movedim.int(arange_3, 0, 0);  arange_3 = None
    select_1: "i64[]" = torch.ops.aten.select.int(movedim_1, 0, 0);  movedim_1 = None
    movedim_2: "i64[s70]" = torch.ops.aten.movedim.int(arange, 0, 0);  arange = movedim_2 = None
    unsqueeze: "i64[1]" = torch.ops.aten.unsqueeze.default(select, 0);  select = None
    expand: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_2]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1]" = torch.ops.aten.unsqueeze.default(select_1, 0);  select_1 = None
    expand_1: "i64[s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_2]);  unsqueeze_1 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_4: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s70, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_2, sym_size_int_4]);  unsqueeze_2 = sym_size_int_2 = sym_size_int_4 = expand_2 = None


cache-cache_patch:torch-rt/inputs_batch1 EXPORT:  78%|███████▊  | 200/256 [00:56<00:23,  2.40it/s]
cache-cache_patch:torch-oblivious/inputs EXPORT:  78%|███████▊  | 200/256 [00:56<00:23,  2.40it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious/inputs EXPORT:  79%|███████▊  | 201/256 [00:57<00:21,  2.59it/s]
cache-cache_patch:torch-oblivious/inputs2 EXPORT:  79%|███████▊  | 201/256 [00:57<00:21,  2.59it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious/inputs2 EXPORT:  79%|███████▉  | 202/256 [00:57<00:20,  2.60it/s]
cache-cache_patch:torch-oblivious/inputs_empty_cache EXPORT:  79%|███████▉  | 202/256 [00:57<00:20,  2.60it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious/inputs_empty_cache EXPORT:  79%|███████▉  | 203/256 [00:57<00:20,  2.61it/s]
cache-cache_patch:torch-oblivious/inputs_batch1 EXPORT:  79%|███████▉  | 203/256 [00:57<00:20,  2.61it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious/inputs_batch1 EXPORT:  80%|███████▉  | 204/256 [00:58<00:22,  2.35it/s]
cache-cache_patch:torch-oblivious-rt/inputs EXPORT:  80%|███████▉  | 204/256 [00:58<00:22,  2.35it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious-rt/inputs EXPORT:  80%|████████  | 205/256 [00:58<00:21,  2.40it/s]
cache-cache_patch:torch-oblivious-rt/inputs2 EXPORT:  80%|████████  | 205/256 [00:58<00:21,  2.40it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious-rt/inputs2 EXPORT:  80%|████████  | 206/256 [00:59<00:19,  2.55it/s]
cache-cache_patch:torch-oblivious-rt/inputs_empty_cache EXPORT:  80%|████████  | 206/256 [00:59<00:19,  2.55it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious-rt/inputs_empty_cache EXPORT:  81%|████████  | 207/256 [00:59<00:17,  2.79it/s]
cache-cache_patch:torch-oblivious-rt/inputs_batch1 EXPORT:  81%|████████  | 207/256 [00:59<00:17,  2.79it/s]


def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None




def forward(self, arg0_1: "f32[32000, 192]", arg1_1: "f32[192, 192]", arg2_1: "f32[96, 192]", arg3_1: "f32[96, 192]", arg4_1: "f32[192, 192]", arg5_1: "f32[1024, 192]", arg6_1: "f32[1024, 192]", arg7_1: "f32[192, 1024]", arg8_1: "f32[192]", arg9_1: "f32[192]", arg10_1: "f32[192]", arg11_1: "f32[32000, 192]", arg12_1: "f32[48]", arg13_1: "i64[s72, s70]", arg14_1: "i64[s43, s53]", arg15_1: "i64[s44, s9]", arg16_1: "f32[s23, 1, s31, 96]", arg17_1: "f32[s4, 1, s11, 96]"):
    # No stacktrace found for following nodes
    _tensor_constant0: "f32[0]" = self._tensor_constant0
    lift_fresh_copy: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant0);  _tensor_constant0 = None
    detach_: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy);  lift_fresh_copy = None
    _tensor_constant1: "f32[0]" = self._tensor_constant1
    lift_fresh_copy_1: "f32[0]" = torch.ops.aten.lift_fresh_copy.default(_tensor_constant1);  _tensor_constant1 = None
    detach__1: "f32[0]" = torch.ops.aten.detach_.default(lift_fresh_copy_1);  lift_fresh_copy_1 = None
    cat: "f32[s23, 1, s31, 96]" = torch.ops.aten.cat.default([detach_, arg16_1], -2);  detach_ = arg16_1 = None
    cat_1: "f32[s4, 1, s11, 96]" = torch.ops.aten.cat.default([detach__1, arg17_1], -2);  detach__1 = arg17_1 = cat_1 = None

     # File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
    embedding: "f32[s72, s70, 192]" = torch.ops.aten.embedding.default(arg0_1, arg13_1);  arg0_1 = embedding = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:371 in forward, code: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
    sym_numel_default: "Sym(96*s23*s31)" = torch.ops.aten.sym_numel.default(cat)
    eq: "Sym(False)" = sym_numel_default == 0;  eq = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:373 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
    sym_size_int: "Sym(s31)" = torch.ops.aten.sym_size.int(cat, 2);  cat = None
    sym_size_int_1: "Sym(s70)" = torch.ops.aten.sym_size.int(arg13_1, 1)
    add: "Sym(s31 + s70)" = sym_size_int + sym_size_int_1;  sym_size_int_1 = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:372 in forward, code: cache_position: torch.Tensor = torch.arange(
    arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int, add, device = device(type='cpu'), pin_memory = False);  add = None

     # File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:379 in forward, code: causal_mask = create_causal_mask(
    to: "b8[s43, s53]" = torch.ops.aten.to.device(arg14_1, device(type='cpu'), torch.bool);  to = None
    eq_1: "Sym(False)" = sym_numel_default == 0;  sym_numel_default = eq_1 = None
    sym_size_int_2: "Sym(s70)" = torch.ops.aten.sym_size.int(arange, 0)
    add_1: "Sym(s31 + s70)" = sym_size_int + sym_size_int_2;  sym_size_int = None
    add_2: "Sym(s31 + s70)" = add_1 + 0
    sym_size_int_3: "Sym(s53)" = torch.ops.aten.sym_size.int(arg14_1, 1);  arg14_1 = None
    sub: "Sym(s31 - s53 + s70)" = add_2 - sym_size_int_3;  add_2 = None
    gt: "Sym(s31 - s53 + s70 > 0)" = sub > 0;  sub = gt = None
    gt_1: "Sym(s53 > s31 + s70)" = sym_size_int_3 > add_1;  sym_size_int_3 = gt_1 = None
    arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add_1, device = device(type='cpu'), pin_memory = False);  add_1 = None
    add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0)
    sym_size_int_4: "Sym(s72)" = torch.ops.aten.sym_size.int(arg13_1, 0);  arg13_1 = None
    arange_2: "i64[s72]" = torch.ops.aten.arange.default(sym_size_int_4, device = device(type='cpu'), pin_memory = False);  sym_size_int_4 = None
    arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
    movedim: "i64[s72]" = torch.ops.aten.movedim.int(arange_2, 0, 0);  movedim = None
    unsqueeze: "i64[1, 1]" = torch.ops.aten.unsqueeze.default(arange_3, 0);  arange_3 = None
    sym_size_int_5: "Sym(s72)" = torch.ops.aten.sym_size.int(arange_2, 0);  arange_2 = None
    expand: "i64[s72, 1]" = torch.ops.aten.expand.default(unsqueeze, [sym_size_int_5, 1]);  unsqueeze = expand = None
    unsqueeze_1: "i64[1, s70]" = torch.ops.aten.unsqueeze.default(arange, 0);  arange = None
    expand_1: "i64[s72, s70]" = torch.ops.aten.expand.default(unsqueeze_1, [sym_size_int_5, sym_size_int_2]);  unsqueeze_1 = sym_size_int_2 = expand_1 = None
    unsqueeze_2: "i64[1, s31 + s70]" = torch.ops.aten.unsqueeze.default(add_, 0);  add_ = None
    sym_size_int_6: "Sym(s31 + s70)" = torch.ops.aten.sym_size.int(arange_1, 0);  arange_1 = None
    expand_2: "i64[s72, s31 + s70]" = torch.ops.aten.expand.default(unsqueeze_2, [sym_size_int_5, sym_size_int_6]);  unsqueeze_2 = sym_size_int_5 = sym_size_int_6 = expand_2 = None


cache-cache_patch:torch-oblivious-rt/inputs_batch1 EXPORT:  81%|████████▏ | 208/256 [00:59<00:16,  2.95it/s]
cache-cache_patch:torch-strict/inputs EXPORT:  81%|████████▏ | 208/256 [00:59<00:16,  2.95it/s]
cache-cache_patch:torch-strict/inputs EXPORT:  82%|████████▏ | 209/256 [00:59<00:14,  3.29it/s]
cache-cache_patch:torch-strict/inputs2 EXPORT:  82%|████████▏ | 209/256 [00:59<00:14,  3.29it/s]
cache-cache_patch:torch-strict/inputs2 EXPORT:  82%|████████▏ | 210/256 [01:00<00:13,  3.49it/s]
cache-cache_patch:torch-strict/inputs_empty_cache EXPORT:  82%|████████▏ | 210/256 [01:00<00:13,  3.49it/s]
cache-cache_patch:torch-strict/inputs_empty_cache EXPORT:  82%|████████▏ | 211/256 [01:00<00:11,  4.08it/s]
cache-cache_patch:torch-strict/inputs_batch1 EXPORT:  82%|████████▏ | 211/256 [01:00<00:11,  4.08it/s]
cache-cache_patch:torch-strict/inputs_batch1 EXPORT:  83%|████████▎ | 212/256 [01:00<00:09,  4.50it/s]
cache-cache_patch:torch-strict-rt/inputs EXPORT:  83%|████████▎ | 212/256 [01:00<00:09,  4.50it/s]
cache-cache_patch:torch-strict-rt/inputs EXPORT:  83%|████████▎ | 213/256 [01:00<00:09,  4.44it/s]
cache-cache_patch:torch-strict-rt/inputs2 EXPORT:  83%|████████▎ | 213/256 [01:00<00:09,  4.44it/s]
cache-cache_patch:torch-strict-rt/inputs2 EXPORT:  84%|████████▎ | 214/256 [01:00<00:09,  4.26it/s]
cache-cache_patch:torch-strict-rt/inputs_empty_cache EXPORT:  84%|████████▎ | 214/256 [01:00<00:09,  4.26it/s]
cache-cache_patch:torch-strict-rt/inputs_empty_cache EXPORT:  84%|████████▍ | 215/256 [01:01<00:08,  4.83it/s]
cache-cache_patch:torch-strict-rt/inputs_batch1 EXPORT:  84%|████████▍ | 215/256 [01:01<00:08,  4.83it/s]
cache-cache_patch:torch-strict-rt/inputs_batch1 EXPORT:  84%|████████▍ | 216/256 [01:01<00:07,  5.42it/s]
cache-cache_patch:torch-strict-oblivious/inputs EXPORT:  84%|████████▍ | 216/256 [01:01<00:07,  5.42it/s]
cache-cache_patch:torch-strict-oblivious/inputs EXPORT:  85%|████████▍ | 217/256 [01:01<00:06,  5.85it/s]
cache-cache_patch:torch-strict-oblivious/inputs2 EXPORT:  85%|████████▍ | 217/256 [01:01<00:06,  5.85it/s]
cache-cache_patch:torch-strict-oblivious/inputs2 EXPORT:  85%|████████▌ | 218/256 [01:01<00:06,  6.25it/s]
cache-cache_patch:torch-strict-oblivious/inputs_empty_cache EXPORT:  85%|████████▌ | 218/256 [01:01<00:06,  6.25it/s]
cache-cache_patch:torch-strict-oblivious/inputs_empty_cache EXPORT:  86%|████████▌ | 219/256 [01:01<00:05,  6.53it/s]
cache-cache_patch:torch-strict-oblivious/inputs_batch1 EXPORT:  86%|████████▌ | 219/256 [01:01<00:05,  6.53it/s]
cache-cache_patch:torch-strict-oblivious/inputs_batch1 EXPORT:  86%|████████▌ | 220/256 [01:01<00:05,  6.65it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs EXPORT:  86%|████████▌ | 220/256 [01:01<00:05,  6.65it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs EXPORT:  86%|████████▋ | 221/256 [01:01<00:05,  6.86it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs2 EXPORT:  86%|████████▋ | 221/256 [01:01<00:05,  6.86it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs2 EXPORT:  87%|████████▋ | 222/256 [01:02<00:04,  6.97it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs_empty_cache EXPORT:  87%|████████▋ | 222/256 [01:02<00:04,  6.97it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs_empty_cache EXPORT:  87%|████████▋ | 223/256 [01:02<00:04,  7.01it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs_batch1 EXPORT:  87%|████████▋ | 223/256 [01:02<00:04,  7.01it/s]
cache-cache_patch:torch-strict-oblivious-rt/inputs_batch1 EXPORT:  88%|████████▊ | 224/256 [01:02<00:04,  7.08it/s]
cache-cache_patch:transformers/inputs EXPORT:  88%|████████▊ | 224/256 [01:02<00:04,  7.08it/s]
cache-cache_patch:transformers/inputs VALIDATE:  88%|████████▊ | 224/256 [01:03<00:04,  7.08it/s]
cache-cache_patch:transformers/inputs VALIDATE:  88%|████████▊ | 225/256 [01:03<00:12,  2.45it/s]
cache-cache_patch:transformers/inputs2 EXPORT:  88%|████████▊ | 225/256 [01:03<00:12,  2.45it/s]
cache-cache_patch:transformers/inputs2 VALIDATE:  88%|████████▊ | 225/256 [01:04<00:12,  2.45it/s]
cache-cache_patch:transformers/inputs2 VALIDATE:  88%|████████▊ | 226/256 [01:04<00:18,  1.59it/s]
cache-cache_patch:transformers/inputs_empty_cache EXPORT:  88%|████████▊ | 226/256 [01:04<00:18,  1.59it/s]
cache-cache_patch:transformers/inputs_empty_cache EXPORT:  89%|████████▊ | 227/256 [01:06<00:28,  1.00it/s]
cache-cache_patch:transformers/inputs_batch1 EXPORT:  89%|████████▊ | 227/256 [01:06<00:28,  1.00it/s]
cache-cache_patch:transformers/inputs_batch1 EXPORT:  89%|████████▉ | 228/256 [01:07<00:26,  1.04it/s]
cache-cache_patch:transformers-rt/inputs EXPORT:  89%|████████▉ | 228/256 [01:07<00:26,  1.04it/s]
cache-cache_patch:transformers-rt/inputs VALIDATE:  89%|████████▉ | 228/256 [01:08<00:26,  1.04it/s]
cache-cache_patch:transformers-rt/inputs VALIDATE:  89%|████████▉ | 229/256 [01:08<00:26,  1.00it/s]
cache-cache_patch:transformers-rt/inputs2 EXPORT:  89%|████████▉ | 229/256 [01:08<00:26,  1.00it/s]
cache-cache_patch:transformers-rt/inputs2 VALIDATE:  89%|████████▉ | 229/256 [01:09<00:26,  1.00it/s]
cache-cache_patch:transformers-rt/inputs2 VALIDATE:  90%|████████▉ | 230/256 [01:09<00:26,  1.01s/it]
cache-cache_patch:transformers-rt/inputs_empty_cache EXPORT:  90%|████████▉ | 230/256 [01:09<00:26,  1.01s/it]
cache-cache_patch:transformers-rt/inputs_empty_cache EXPORT:  90%|█████████ | 231/256 [01:10<00:24,  1.01it/s]
cache-cache_patch:transformers-rt/inputs_batch1 EXPORT:  90%|█████████ | 231/256 [01:10<00:24,  1.01it/s]
cache-cache_patch:transformers-rt/inputs_batch1 EXPORT:  91%|█████████ | 232/256 [01:11<00:22,  1.09it/s]
cache-cache_patch:transformers-oblivious/inputs EXPORT:  91%|█████████ | 232/256 [01:11<00:22,  1.09it/s]
cache-cache_patch:transformers-oblivious/inputs VALIDATE:  91%|█████████ | 232/256 [01:10<00:22,  1.09it/s]
cache-cache_patch:transformers-oblivious/inputs2 EXPORT:  91%|█████████ | 232/256 [01:10<00:22,  1.09it/s]
cache-cache_patch:transformers-oblivious/inputs2 VALIDATE:  91%|█████████ | 232/256 [01:11<00:22,  1.09it/s]
cache-cache_patch:transformers-oblivious/inputs2 VALIDATE:  91%|█████████▏| 234/256 [01:11<00:13,  1.59it/s]
cache-cache_patch:transformers-oblivious/inputs_empty_cache EXPORT:  91%|█████████▏| 234/256 [01:11<00:13,  1.59it/s]
cache-cache_patch:transformers-oblivious/inputs_empty_cache VALIDATE:  91%|█████████▏| 234/256 [01:12<00:13,  1.59it/s]
cache-cache_patch:transformers-oblivious/inputs_empty_cache VALIDATE:  92%|█████████▏| 235/256 [01:12<00:14,  1.41it/s]
cache-cache_patch:transformers-oblivious/inputs_batch1 EXPORT:  92%|█████████▏| 235/256 [01:12<00:14,  1.41it/s]
cache-cache_patch:transformers-oblivious/inputs_batch1 VALIDATE:  92%|█████████▏| 235/256 [01:13<00:14,  1.41it/s]
cache-cache_patch:transformers-oblivious/inputs_batch1 VALIDATE:  92%|█████████▏| 236/256 [01:13<00:15,  1.31it/s]
cache-cache_patch:transformers-oblivious-rt/inputs EXPORT:  92%|█████████▏| 236/256 [01:13<00:15,  1.31it/s]
cache-cache_patch:transformers-oblivious-rt/inputs VALIDATE:  92%|█████████▏| 236/256 [01:14<00:15,  1.31it/s]
cache-cache_patch:transformers-oblivious-rt/inputs VALIDATE:  93%|█████████▎| 237/256 [01:14<00:16,  1.16it/s]
cache-cache_patch:transformers-oblivious-rt/inputs2 EXPORT:  93%|█████████▎| 237/256 [01:14<00:16,  1.16it/s]
cache-cache_patch:transformers-oblivious-rt/inputs2 VALIDATE:  93%|█████████▎| 237/256 [01:15<00:16,  1.16it/s]
cache-cache_patch:transformers-oblivious-rt/inputs2 VALIDATE:  93%|█████████▎| 238/256 [01:15<00:16,  1.11it/s]
cache-cache_patch:transformers-oblivious-rt/inputs_empty_cache EXPORT:  93%|█████████▎| 238/256 [01:15<00:16,  1.11it/s]
cache-cache_patch:transformers-oblivious-rt/inputs_empty_cache VALIDATE:  93%|█████████▎| 238/256 [01:16<00:16,  1.11it/s]
cache-cache_patch:transformers-oblivious-rt/inputs_empty_cache VALIDATE:  93%|█████████▎| 239/256 [01:16<00:15,  1.08it/s]
cache-cache_patch:transformers-oblivious-rt/inputs_batch1 EXPORT:  93%|█████████▎| 239/256 [01:16<00:15,  1.08it/s]
cache-cache_patch:transformers-oblivious-rt/inputs_batch1 VALIDATE:  93%|█████████▎| 239/256 [01:17<00:15,  1.08it/s]
cache-cache_patch:transformers-oblivious-rt/inputs_batch1 VALIDATE:  94%|█████████▍| 240/256 [01:17<00:14,  1.09it/s]
cache-cache_patch:transformers-strict/inputs EXPORT:  94%|█████████▍| 240/256 [01:17<00:14,  1.09it/s]               ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict/inputs EXPORT:  94%|█████████▍| 241/256 [01:18<00:14,  1.06it/s]
cache-cache_patch:transformers-strict/inputs2 EXPORT:  94%|█████████▍| 241/256 [01:18<00:14,  1.06it/s]
cache-cache_patch:transformers-strict/inputs2 EXPORT:  95%|█████████▍| 242/256 [01:20<00:17,  1.27s/it]
cache-cache_patch:transformers-strict/inputs_empty_cache EXPORT:  95%|█████████▍| 242/256 [01:20<00:17,  1.27s/it]
cache-cache_patch:transformers-strict/inputs_empty_cache EXPORT:  95%|█████████▍| 243/256 [01:21<00:15,  1.20s/it]
cache-cache_patch:transformers-strict/inputs_batch1 EXPORT:  95%|█████████▍| 243/256 [01:21<00:15,  1.20s/it]
cache-cache_patch:transformers-strict/inputs_batch1 EXPORT:  95%|█████████▌| 244/256 [01:22<00:14,  1.23s/it]
cache-cache_patch:transformers-strict-rt/inputs EXPORT:  95%|█████████▌| 244/256 [01:22<00:14,  1.23s/it]    ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-rt/inputs EXPORT:  96%|█████████▌| 245/256 [01:24<00:15,  1.43s/it]
cache-cache_patch:transformers-strict-rt/inputs2 EXPORT:  96%|█████████▌| 245/256 [01:24<00:15,  1.43s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-rt/inputs2 EXPORT:  96%|█████████▌| 246/256 [01:26<00:15,  1.55s/it]
cache-cache_patch:transformers-strict-rt/inputs_empty_cache EXPORT:  96%|█████████▌| 246/256 [01:26<00:15,  1.55s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-rt/inputs_empty_cache EXPORT:  96%|█████████▋| 247/256 [01:28<00:15,  1.70s/it]
cache-cache_patch:transformers-strict-rt/inputs_batch1 EXPORT:  96%|█████████▋| 247/256 [01:28<00:15,  1.70s/it]     ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-rt/inputs_batch1 EXPORT:  97%|█████████▋| 248/256 [01:30<00:12,  1.61s/it]
cache-cache_patch:transformers-strict-oblivious/inputs EXPORT:  97%|█████████▋| 248/256 [01:30<00:12,  1.61s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-oblivious/inputs EXPORT:  97%|█████████▋| 249/256 [01:31<00:09,  1.42s/it]
cache-cache_patch:transformers-strict-oblivious/inputs2 EXPORT:  97%|█████████▋| 249/256 [01:31<00:09,  1.42s/it]
cache-cache_patch:transformers-strict-oblivious/inputs2 EXPORT:  98%|█████████▊| 250/256 [01:31<00:07,  1.26s/it]
cache-cache_patch:transformers-strict-oblivious/inputs_empty_cache EXPORT:  98%|█████████▊| 250/256 [01:31<00:07,  1.26s/it]
cache-cache_patch:transformers-strict-oblivious/inputs_empty_cache EXPORT:  98%|█████████▊| 251/256 [01:32<00:05,  1.14s/it]
cache-cache_patch:transformers-strict-oblivious/inputs_batch1 EXPORT:  98%|█████████▊| 251/256 [01:32<00:05,  1.14s/it]
cache-cache_patch:transformers-strict-oblivious/inputs_batch1 EXPORT:  98%|█████████▊| 252/256 [01:35<00:06,  1.58s/it]
cache-cache_patch:transformers-strict-oblivious-rt/inputs EXPORT:  98%|█████████▊| 252/256 [01:35<00:06,  1.58s/it]    ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-oblivious-rt/inputs EXPORT:  99%|█████████▉| 253/256 [01:37<00:04,  1.64s/it]
cache-cache_patch:transformers-strict-oblivious-rt/inputs2 EXPORT:  99%|█████████▉| 253/256 [01:37<00:04,  1.64s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-oblivious-rt/inputs2 EXPORT:  99%|█████████▉| 254/256 [01:39<00:03,  1.70s/it]
cache-cache_patch:transformers-strict-oblivious-rt/inputs_empty_cache EXPORT:  99%|█████████▉| 254/256 [01:39<00:03,  1.70s/it]~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-oblivious-rt/inputs_empty_cache EXPORT: 100%|█████████▉| 255/256 [01:40<00:01,  1.76s/it]
cache-cache_patch:transformers-strict-oblivious-rt/inputs_batch1 EXPORT: 100%|█████████▉| 255/256 [01:40<00:01,  1.76s/it]     ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1824: UserWarning: While exporting, we found certain side effects happened in the model.forward. Here are the list of potential sources you can double check: ["L['kwargs']['past_key_values'].layers[0]"]
  warnings.warn(

cache-cache_patch:transformers-strict-oblivious-rt/inputs_batch1 EXPORT: 100%|██████████| 256/256 [01:40<00:00,  2.54it/s]

Let’s save the results.

df = pandas.DataFrame(results)
df.to_excel("plot_export_tiny_llm_dim01.xlsx")
df
cache cache_patch strict oblivious rt export_with EXPORT ERR-EXPORT run_with WORKS ERR-RUN
0 0 0 0 0 0 inputs 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
1 0 0 0 0 0 inputs2 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
2 0 0 0 0 0 inputs_empty_cache 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
3 0 0 0 0 0 inputs_batch1 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
4 0 0 0 0 1 inputs 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
... ... ... ... ... ... ... ... ... ... ... ...
239 1 transformers 1 1 0 inputs_batch1 0 Found the following conflicts between user-spe... NaN NaN NaN
240 1 transformers 1 1 1 inputs 0 Found the following conflicts between user-spe... NaN NaN NaN
241 1 transformers 1 1 1 inputs2 0 Found the following conflicts between user-spe... NaN NaN NaN
242 1 transformers 1 1 1 inputs_empty_cache 0 Found the following conflicts between user-spe... NaN NaN NaN
243 1 transformers 1 1 1 inputs_batch1 0 Found the following conflicts between user-spe... NaN NaN NaN

244 rows × 11 columns



no_export = df[df.EXPORT == 0]
no_export.to_excel("plot_export_tiny_llm_dim01.no_export.xlsx")
no_export
cache cache_patch strict oblivious rt export_with EXPORT ERR-EXPORT run_with WORKS ERR-RUN
0 0 0 0 0 0 inputs 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
1 0 0 0 0 0 inputs2 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
2 0 0 0 0 0 inputs_empty_cache 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
3 0 0 0 0 0 inputs_batch1 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
4 0 0 0 0 1 inputs 0 Cannot associate shape [[{0: DimHint(DYNAMIC),... NaN NaN NaN
... ... ... ... ... ... ... ... ... ... ... ...
239 1 transformers 1 1 0 inputs_batch1 0 Found the following conflicts between user-spe... NaN NaN NaN
240 1 transformers 1 1 1 inputs 0 Found the following conflicts between user-spe... NaN NaN NaN
241 1 transformers 1 1 1 inputs2 0 Found the following conflicts between user-spe... NaN NaN NaN
242 1 transformers 1 1 1 inputs_empty_cache 0 Found the following conflicts between user-spe... NaN NaN NaN
243 1 transformers 1 1 1 inputs_batch1 0 Found the following conflicts between user-spe... NaN NaN NaN

132 rows × 11 columns



The validation failures.

invalid = df[(df.EXPORT == 1) & (df.WORKS == 0)].pivot(
    index=["cache", "cache_patch", "strict", "oblivious", "rt", "export_with"],
    columns=["run_with"],
    values=["WORKS", "ERR-RUN"],
)
invalid.to_excel("plot_export_tiny_llm_dim01.invalid.xlsx")
invalid
WORKS ERR-RUN
run_with inputs inputs2 inputs_batch1 inputs_empty_cache inputs inputs2 inputs_batch1 inputs_empty_cache
cache cache_patch strict oblivious rt export_with
1 all 0 0 0 inputs NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
inputs2 NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
inputs_batch1 0.0 0.0 NaN 0.0 Guard failed: input_ids.size()[0] == 1 Guard failed: input_ids.size()[0] == 1 NaN Guard failed: input_ids.size()[0] == 1
inputs_empty_cache 0.0 0.0 0.0 NaN Guard failed: past_key_values['key_cache'][0].... Guard failed: past_key_values['key_cache'][0].... Guard failed: past_key_values['key_cache'][0].... NaN
1 inputs NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
inputs2 NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
inputs_batch1 0.0 0.0 NaN 0.0 Guard failed: input_ids.size()[0] == 1 Guard failed: input_ids.size()[0] == 1 NaN Guard failed: input_ids.size()[0] == 1
inputs_empty_cache 0.0 0.0 0.0 NaN Guard failed: past_key_values['key_cache'][0].... Guard failed: past_key_values['key_cache'][0].... Guard failed: past_key_values['key_cache'][0].... NaN
1 0 inputs NaN NaN 0.0 NaN NaN NaN Guard failed: position_ids.size()[0] != 1 NaN
inputs2 NaN NaN 0.0 NaN NaN NaN Guard failed: position_ids.size()[0] != 1 NaN
inputs_batch1 0.0 0.0 NaN 0.0 Guard failed: position_ids.size()[0] == 1 Guard failed: input_ids.size()[0] <= 2 NaN Guard failed: position_ids.size()[0] == 1
inputs_empty_cache NaN NaN 0.0 NaN NaN NaN Guard failed: position_ids.size()[0] != 1 NaN
1 inputs NaN NaN 0.0 NaN NaN NaN Runtime assertion failed for expression Ne(s44... NaN
inputs2 NaN NaN 0.0 NaN NaN NaN Runtime assertion failed for expression Ne(s44... NaN
inputs_batch1 0.0 0.0 NaN 0.0 Runtime assertion failed for expression Eq(s44... Guard failed: input_ids.size()[0] <= 2 NaN Runtime assertion failed for expression Eq(s44...
inputs_empty_cache NaN NaN 0.0 NaN NaN NaN Runtime assertion failed for expression Ne(s44... NaN
transformers 0 0 0 inputs NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
inputs2 NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
1 inputs NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
inputs2 NaN NaN NaN 0.0 NaN NaN NaN Guard failed: attention_mask.size()[1] >= 4
1 0 inputs NaN NaN 0.0 NaN NaN NaN Guard failed: position_ids.size()[0] != 1 NaN
inputs2 NaN NaN 0.0 NaN NaN NaN Guard failed: position_ids.size()[0] != 1 NaN
inputs_batch1 0.0 0.0 NaN 0.0 Guard failed: position_ids.size()[0] == 1 Guard failed: input_ids.size()[0] <= 2 NaN Guard failed: position_ids.size()[0] == 1
inputs_empty_cache NaN NaN 0.0 NaN NaN NaN Guard failed: position_ids.size()[0] != 1 NaN
1 inputs NaN NaN 0.0 NaN NaN NaN Runtime assertion failed for expression Ne(s44... NaN
inputs2 NaN NaN 0.0 NaN NaN NaN Runtime assertion failed for expression Ne(s44... NaN
inputs_batch1 0.0 0.0 NaN 0.0 Runtime assertion failed for expression Eq(s44... Guard failed: input_ids.size()[0] <= 2 NaN Runtime assertion failed for expression Eq(s44...
inputs_empty_cache NaN NaN 0.0 NaN NaN NaN Runtime assertion failed for expression Ne(s44... NaN 


success = df[(df.EXPORT == 1) & (df.WORKS == 1)].pivot(
    index=["cache", "cache_patch", "strict", "oblivious", "rt", "export_with"],
    columns=["run_with"],
    values=["WORKS"],
)
success.to_excel("plot_export_tiny_llm_dim01.success.xlsx")
success
WORKS
run_with inputs inputs2 inputs_batch1 inputs_empty_cache
cache cache_patch strict oblivious rt export_with
1 all 0 0 0 inputs 1.0 1.0 1.0 NaN
inputs2 1.0 1.0 1.0 NaN
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache NaN NaN NaN 1.0
1 inputs 1.0 1.0 1.0 NaN
inputs2 1.0 1.0 1.0 NaN
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache NaN NaN NaN 1.0
1 0 inputs 1.0 1.0 NaN 1.0
inputs2 1.0 1.0 NaN 1.0
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache 1.0 1.0 NaN 1.0
1 inputs 1.0 1.0 NaN 1.0
inputs2 1.0 1.0 NaN 1.0
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache 1.0 1.0 NaN 1.0
transformers 0 0 0 inputs 1.0 1.0 1.0 NaN
inputs2 1.0 1.0 1.0 NaN
1 inputs 1.0 1.0 1.0 NaN
inputs2 1.0 1.0 1.0 NaN
1 0 inputs 1.0 1.0 NaN 1.0
inputs2 1.0 1.0 NaN 1.0
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache 1.0 1.0 NaN 1.0
1 inputs 1.0 1.0 NaN 1.0
inputs2 1.0 1.0 NaN 1.0
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache 1.0 1.0 NaN 1.0


If you have any error, then look at example Export Tiny-LLM with patches.

doc.plot_legend("Tiny-LLM\nexport with\ndimension in {0,1}", "torch.export.export", "tomato")
plot export tiny llm dim01

Total running time of the script: (1 minutes 47.054 seconds)

Related examples

Export with dynamic dimensions in {0,1} into ONNX

Export with dynamic dimensions in {0,1} into ONNX

Export with dynamic dimensions in {0,1} into ONNX (custom)

Export with dynamic dimensions in {0,1} into ONNX (custom)

Export microsoft/phi-2

Export microsoft/phi-2

Gallery generated by Sphinx-Gallery