Note

Go to the end to download the full example code.

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

This duplicates the example Export with dynamic dimensions in {0,1} but for torch.onnx.export(). It checks what inputs can be used to export and with which inputs it can work.

Available input sets

import itertools
from tqdm import tqdm
import numpy as np
import pandas
import torch
import onnxruntime
from onnx_diagnostic import doc
from onnx_diagnostic.helpers import max_diff, string_type, flatten_object
from onnx_diagnostic.helpers.torch_helper import torch_deepcopy
from onnx_diagnostic.helpers.rt_helper import make_feeds
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)}]])

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
):
    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)
    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)
    if oblivious:
        with torch.fx.experimental._config.patch(backed_size_oblivious=True):
            return export_model(model, dynamic_shapes, inputs, rt=rt)
    ep = torch.export.export(
        model,
        (),
        inputs,
        dynamic_shapes=use_dyn_not_str(dynamic_shapes),
        prefer_deferred_runtime_asserts_over_guards=rt,
    )
    return torch.onnx.export(ep, args=(), kwargs=inputs, dynamic_shapes=dynamic_shapes)


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


def validation(ep, input_sets, expected, catch_exception=True):
    sess = onnxruntime.InferenceSession(
        ep.model_proto.SerializeToString(), providers=["CPUExecutionProvider"]
    )
    for k, v in input_sets.items():
        feeds = make_feeds(sess, torch_deepcopy(v), use_numpy=True)
        try:
            got = sess.run(None, feeds)
        except Exception as e:
            if not catch_exception:
                raise
            yield k, e
            continue
        yield k, max_diff(flatten_object(expected[k], drop_keys=True), got)

Verification an example known to be working is.

ep = export_model(
    model,
    dynamic_shapes,
    torch_deepcopy(input_sets["inputs"]),
    cache_patch=True,
)
res = list(validation(ep, dict(inputs=input_sets["inputs"]), expected, catch_exception=False))
assert res[0][1]["abs"] < 1e-5, f"Unexpected issue with res={res}"

[torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

The main loop

results = []

possibilities = [*[[0, 1] for _ in range(4)], list(input_sets)]
possibilities[1] = [0, "all", "torch", "transformers"]
with tqdm(list(itertools.product(*possibilities))) as pbar:
    for cache, cache_patch, oblivious, rt, inputs in pbar:
        if cache_patch and not cache:
            # patches include caches.
            continue
        kwargs = dict(cache=cache, cache_patch=cache_patch, 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%|          | 0/128 [00:00<?, ?it/s]
/inputs EXPORT:   0%|          | 0/128 [00:00<?, ?it/s]
/inputs EXPORT:   1%|          | 1/128 [00:00<00:24,  5.11it/s]
/inputs2 EXPORT:   1%|          | 1/128 [00:00<00:24,  5.11it/s]
/inputs2 EXPORT:   2%|▏         | 2/128 [00:00<00:23,  5.35it/s]
/inputs_empty_cache EXPORT:   2%|▏         | 2/128 [00:00<00:23,  5.35it/s]
/inputs_empty_cache EXPORT:   2%|▏         | 3/128 [00:00<00:21,  5.77it/s]
/inputs_batch1 EXPORT:   2%|▏         | 3/128 [00:00<00:21,  5.77it/s]
/inputs_batch1 EXPORT:   3%|▎         | 4/128 [00:00<00:23,  5.23it/s]
rt/inputs EXPORT:   3%|▎         | 4/128 [00:00<00:23,  5.23it/s]
rt/inputs EXPORT:   4%|▍         | 5/128 [00:00<00:23,  5.33it/s]
rt/inputs2 EXPORT:   4%|▍         | 5/128 [00:00<00:23,  5.33it/s]
rt/inputs2 EXPORT:   5%|▍         | 6/128 [00:01<00:22,  5.40it/s]
rt/inputs_empty_cache EXPORT:   5%|▍         | 6/128 [00:01<00:22,  5.40it/s]
rt/inputs_empty_cache EXPORT:   5%|▌         | 7/128 [00:01<00:21,  5.66it/s]
rt/inputs_batch1 EXPORT:   5%|▌         | 7/128 [00:01<00:21,  5.66it/s]
rt/inputs_batch1 EXPORT:   6%|▋         | 8/128 [00:01<00:23,  5.09it/s]
oblivious/inputs EXPORT:   6%|▋         | 8/128 [00:01<00:23,  5.09it/s]
oblivious/inputs EXPORT:   7%|▋         | 9/128 [00:01<00:24,  4.89it/s]
oblivious/inputs2 EXPORT:   7%|▋         | 9/128 [00:01<00:24,  4.89it/s]
oblivious/inputs2 EXPORT:   8%|▊         | 10/128 [00:01<00:23,  4.96it/s]
oblivious/inputs_empty_cache EXPORT:   8%|▊         | 10/128 [00:01<00:23,  4.96it/s]
oblivious/inputs_empty_cache EXPORT:   9%|▊         | 11/128 [00:02<00:23,  5.01it/s]
oblivious/inputs_batch1 EXPORT:   9%|▊         | 11/128 [00:02<00:23,  5.01it/s]
oblivious/inputs_batch1 EXPORT:   9%|▉         | 12/128 [00:02<00:22,  5.13it/s]
oblivious-rt/inputs EXPORT:   9%|▉         | 12/128 [00:02<00:22,  5.13it/s]
oblivious-rt/inputs EXPORT:  10%|█         | 13/128 [00:02<00:21,  5.24it/s]
oblivious-rt/inputs2 EXPORT:  10%|█         | 13/128 [00:02<00:21,  5.24it/s]
oblivious-rt/inputs2 EXPORT:  11%|█         | 14/128 [00:02<00:21,  5.28it/s]
oblivious-rt/inputs_empty_cache EXPORT:  11%|█         | 14/128 [00:02<00:21,  5.28it/s]
oblivious-rt/inputs_empty_cache EXPORT:  12%|█▏        | 15/128 [00:02<00:21,  5.26it/s]
oblivious-rt/inputs_batch1 EXPORT:  12%|█▏        | 15/128 [00:02<00:21,  5.26it/s]
oblivious-rt/inputs_batch1 EXPORT:  12%|█▎        | 16/128 [00:03<00:21,  5.14it/s]
cache/inputs EXPORT:  12%|█▎        | 16/128 [00:03<00:21,  5.14it/s]
cache/inputs EXPORT:  51%|█████     | 65/128 [00:03<00:00, 80.44it/s]
cache/inputs2 EXPORT:  51%|█████     | 65/128 [00:03<00:00, 80.44it/s]
cache/inputs_empty_cache EXPORT:  51%|█████     | 65/128 [00:03<00:00, 80.44it/s]
cache/inputs_batch1 EXPORT:  51%|█████     | 65/128 [00:03<00:00, 80.44it/s]
cache-rt/inputs EXPORT:  51%|█████     | 65/128 [00:03<00:00, 80.44it/s]
cache-rt/inputs2 EXPORT:  51%|█████     | 65/128 [00:04<00:00, 80.44it/s]
cache-rt/inputs_empty_cache EXPORT:  51%|█████     | 65/128 [00:04<00:00, 80.44it/s]
cache-rt/inputs_batch1 EXPORT:  51%|█████     | 65/128 [00:04<00:00, 80.44it/s]
cache-oblivious/inputs EXPORT:  51%|█████     | 65/128 [00:04<00:00, 80.44it/s]
cache-oblivious/inputs2 EXPORT:  51%|█████     | 65/128 [00:04<00:00, 80.44it/s]
cache-oblivious/inputs2 EXPORT:  58%|█████▊    | 74/128 [00:04<00:02, 20.86it/s]
cache-oblivious/inputs_empty_cache EXPORT:  58%|█████▊    | 74/128 [00:04<00:02, 20.86it/s]
cache-oblivious/inputs_batch1 EXPORT:  58%|█████▊    | 74/128 [00:05<00:02, 20.86it/s]
cache-oblivious-rt/inputs EXPORT:  58%|█████▊    | 74/128 [00:05<00:02, 20.86it/s]
cache-oblivious-rt/inputs2 EXPORT:  58%|█████▊    | 74/128 [00:05<00:02, 20.86it/s]
cache-oblivious-rt/inputs_empty_cache EXPORT:  58%|█████▊    | 74/128 [00:06<00:02, 20.86it/s]
cache-oblivious-rt/inputs_batch1 EXPORT:  58%|█████▊    | 74/128 [00:07<00:02, 20.86it/s]
cache-oblivious-rt/inputs_batch1 EXPORT:  62%|██████▎   | 80/128 [00:07<00:04,  9.65it/s]
cache-cache_patch:all/inputs EXPORT:  62%|██████▎   | 80/128 [00:07<00:04,  9.65it/s]    [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all/inputs VALIDATE:  62%|██████▎   | 80/128 [00:11<00:04,  9.65it/s]
cache-cache_patch:all/inputs2 EXPORT:  62%|██████▎   | 80/128 [00:11<00:04,  9.65it/s] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all/inputs2 VALIDATE:  62%|██████▎   | 80/128 [00:15<00:04,  9.65it/s]
cache-cache_patch:all/inputs_empty_cache EXPORT:  62%|██████▎   | 80/128 [00:15<00:04,  9.65it/s][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 47 of general pattern rewrite rules.

cache-cache_patch:all/inputs_empty_cache VALIDATE:  62%|██████▎   | 80/128 [00:20<00:04,  9.65it/s]
cache-cache_patch:all/inputs_batch1 EXPORT:  62%|██████▎   | 80/128 [00:20<00:04,  9.65it/s]       [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 17 of general pattern rewrite rules.

cache-cache_patch:all/inputs_batch1 VALIDATE:  62%|██████▎   | 80/128 [00:24<00:04,  9.65it/s]
cache-cache_patch:all-rt/inputs EXPORT:  62%|██████▎   | 80/128 [00:24<00:04,  9.65it/s]      [torch.onnx] Run decomposition...

cache-cache_patch:all-rt/inputs EXPORT:  62%|██████▎   | 80/128 [00:25<00:04,  9.65it/s][torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-rt/inputs VALIDATE:  62%|██████▎   | 80/128 [00:26<00:04,  9.65it/s]
cache-cache_patch:all-rt/inputs VALIDATE:  66%|██████▋   | 85/128 [00:27<00:30,  1.39it/s]
cache-cache_patch:all-rt/inputs2 EXPORT:  66%|██████▋   | 85/128 [00:27<00:30,  1.39it/s] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-rt/inputs2 VALIDATE:  66%|██████▋   | 85/128 [00:31<00:30,  1.39it/s]
cache-cache_patch:all-rt/inputs2 VALIDATE:  67%|██████▋   | 86/128 [00:31<00:37,  1.13it/s]
cache-cache_patch:all-rt/inputs_empty_cache EXPORT:  67%|██████▋   | 86/128 [00:31<00:37,  1.13it/s][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 47 of general pattern rewrite rules.

cache-cache_patch:all-rt/inputs_empty_cache VALIDATE:  67%|██████▋   | 86/128 [00:36<00:37,  1.13it/s]
cache-cache_patch:all-rt/inputs_batch1 EXPORT:  67%|██████▋   | 86/128 [00:37<00:37,  1.13it/s]       [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 17 of general pattern rewrite rules.

cache-cache_patch:all-rt/inputs_batch1 VALIDATE:  67%|██████▋   | 86/128 [00:40<00:37,  1.13it/s]
cache-cache_patch:all-oblivious/inputs EXPORT:  67%|██████▋   | 86/128 [00:40<00:37,  1.13it/s]  [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-oblivious/inputs VALIDATE:  67%|██████▋   | 86/128 [00:44<00:37,  1.13it/s]
cache-cache_patch:all-oblivious/inputs VALIDATE:  70%|██████▉   | 89/128 [00:44<00:56,  1.44s/it]
cache-cache_patch:all-oblivious/inputs2 EXPORT:  70%|██████▉   | 89/128 [00:44<00:56,  1.44s/it] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-oblivious/inputs2 VALIDATE:  70%|██████▉   | 89/128 [00:50<00:56,  1.44s/it]
cache-cache_patch:all-oblivious/inputs2 VALIDATE:  70%|███████   | 90/128 [00:50<01:05,  1.73s/it]
cache-cache_patch:all-oblivious/inputs_empty_cache EXPORT:  70%|███████   | 90/128 [00:50<01:05,  1.73s/it][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-oblivious/inputs_empty_cache VALIDATE:  70%|███████   | 90/128 [00:54<01:05,  1.73s/it]
cache-cache_patch:all-oblivious/inputs_batch1 EXPORT:  70%|███████   | 90/128 [00:54<01:05,  1.73s/it]       [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ❌

cache-cache_patch:all-oblivious-rt/inputs EXPORT:  70%|███████   | 90/128 [00:56<01:05,  1.73s/it]    [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-oblivious-rt/inputs VALIDATE:  70%|███████   | 90/128 [00:58<01:05,  1.73s/it]
cache-cache_patch:all-oblivious-rt/inputs VALIDATE:  73%|███████▎  | 93/128 [00:59<01:10,  2.02s/it]
cache-cache_patch:all-oblivious-rt/inputs2 EXPORT:  73%|███████▎  | 93/128 [00:59<01:10,  2.02s/it] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-oblivious-rt/inputs2 VALIDATE:  73%|███████▎  | 93/128 [01:04<01:10,  2.02s/it]
cache-cache_patch:all-oblivious-rt/inputs_empty_cache EXPORT:  73%|███████▎  | 93/128 [01:05<01:10,  2.02s/it][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:all-oblivious-rt/inputs_empty_cache VALIDATE:  73%|███████▎  | 93/128 [01:09<01:10,  2.02s/it]
cache-cache_patch:all-oblivious-rt/inputs_empty_cache VALIDATE:  74%|███████▍  | 95/128 [01:09<01:25,  2.60s/it]
cache-cache_patch:all-oblivious-rt/inputs_batch1 EXPORT:  74%|███████▍  | 95/128 [01:09<01:25,  2.60s/it]       [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ❌

cache-cache_patch:all-oblivious-rt/inputs_batch1 EXPORT:  75%|███████▌  | 96/128 [01:13<01:29,  2.78s/it]
cache-cache_patch:torch/inputs EXPORT:  75%|███████▌  | 96/128 [01:13<01:29,  2.78s/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:  75%|███████▌  | 96/128 [01:14<01:29,  2.78s/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:  77%|███████▋  | 98/128 [01:14<01:04,  2.16s/it]
cache-cache_patch:torch/inputs_empty_cache EXPORT:  77%|███████▋  | 98/128 [01:14<01:04,  2.16s/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:  77%|███████▋  | 99/128 [01:14<00:54,  1.87s/it]
cache-cache_patch:torch/inputs_batch1 EXPORT:  77%|███████▋  | 99/128 [01:14<00:54,  1.87s/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:  78%|███████▊  | 100/128 [01:14<00:43,  1.57s/it]
cache-cache_patch:torch-rt/inputs EXPORT:  78%|███████▊  | 100/128 [01:14<00:43,  1.57s/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-rt/inputs EXPORT:  79%|███████▉  | 101/128 [01:15<00:35,  1.30s/it]
cache-cache_patch:torch-rt/inputs2 EXPORT:  79%|███████▉  | 101/128 [01:15<00:35,  1.30s/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-rt/inputs2 EXPORT:  80%|███████▉  | 102/128 [01:15<00:27,  1.05s/it]
cache-cache_patch:torch-rt/inputs_empty_cache EXPORT:  80%|███████▉  | 102/128 [01:15<00:27,  1.05s/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-rt/inputs_empty_cache EXPORT:  80%|████████  | 103/128 [01:15<00:21,  1.18it/s]
cache-cache_patch:torch-rt/inputs_batch1 EXPORT:  80%|████████  | 103/128 [01:15<00:21,  1.18it/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:  81%|████████▏ | 104/128 [01:15<00:16,  1.46it/s]
cache-cache_patch:torch-oblivious/inputs EXPORT:  81%|████████▏ | 104/128 [01:15<00:16,  1.46it/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:  82%|████████▏ | 105/128 [01:15<00:12,  1.79it/s]
cache-cache_patch:torch-oblivious/inputs2 EXPORT:  82%|████████▏ | 105/128 [01:15<00:12,  1.79it/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/inputs2 EXPORT:  83%|████████▎ | 106/128 [01:16<00:10,  2.16it/s]
cache-cache_patch:torch-oblivious/inputs_empty_cache EXPORT:  83%|████████▎ | 106/128 [01:16<00:10,  2.16it/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:  84%|████████▎ | 107/128 [01:16<00:08,  2.52it/s]
cache-cache_patch:torch-oblivious/inputs_batch1 EXPORT:  84%|████████▎ | 107/128 [01:16<00:08,  2.52it/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:  84%|████████▍ | 108/128 [01:16<00:07,  2.84it/s]
cache-cache_patch:torch-oblivious-rt/inputs EXPORT:  84%|████████▍ | 108/128 [01:16<00:07,  2.84it/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:  85%|████████▌ | 109/128 [01:16<00:06,  3.07it/s]
cache-cache_patch:torch-oblivious-rt/inputs2 EXPORT:  85%|████████▌ | 109/128 [01:16<00:06,  3.07it/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:  86%|████████▌ | 110/128 [01:17<00:05,  3.35it/s]
cache-cache_patch:torch-oblivious-rt/inputs_empty_cache EXPORT:  86%|████████▌ | 110/128 [01:17<00:05,  3.35it/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_empty_cache EXPORT:  87%|████████▋ | 111/128 [01:17<00:04,  3.56it/s]
cache-cache_patch:torch-oblivious-rt/inputs_batch1 EXPORT:  87%|████████▋ | 111/128 [01:17<00:04,  3.56it/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:  88%|████████▊ | 112/128 [01:17<00:04,  3.80it/s]
cache-cache_patch:transformers/inputs EXPORT:  88%|████████▊ | 112/128 [01:17<00:04,  3.80it/s]             [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers/inputs VALIDATE:  88%|████████▊ | 112/128 [01:23<00:04,  3.80it/s]
cache-cache_patch:transformers/inputs VALIDATE:  88%|████████▊ | 113/128 [01:23<00:29,  1.94s/it]
cache-cache_patch:transformers/inputs2 EXPORT:  88%|████████▊ | 113/128 [01:23<00:29,  1.94s/it] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers/inputs2 VALIDATE:  88%|████████▊ | 113/128 [01:25<00:29,  1.94s/it]
cache-cache_patch:transformers/inputs2 VALIDATE:  89%|████████▉ | 114/128 [01:26<00:30,  2.16s/it]
cache-cache_patch:transformers/inputs_empty_cache EXPORT:  89%|████████▉ | 114/128 [01:26<00:30,  2.16s/it]
cache-cache_patch:transformers/inputs_empty_cache EXPORT:  90%|████████▉ | 115/128 [01:27<00:23,  1.78s/it]
cache-cache_patch:transformers/inputs_batch1 EXPORT:  90%|████████▉ | 115/128 [01:27<00:23,  1.78s/it]
cache-cache_patch:transformers/inputs_batch1 EXPORT:  91%|█████████ | 116/128 [01:27<00:17,  1.46s/it]
cache-cache_patch:transformers-rt/inputs EXPORT:  91%|█████████ | 116/128 [01:27<00:17,  1.46s/it]    [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-rt/inputs VALIDATE:  91%|█████████ | 116/128 [01:31<00:17,  1.46s/it]
cache-cache_patch:transformers-rt/inputs VALIDATE:  91%|█████████▏| 117/128 [01:31<00:25,  2.30s/it]
cache-cache_patch:transformers-rt/inputs2 EXPORT:  91%|█████████▏| 117/128 [01:31<00:25,  2.30s/it] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-rt/inputs2 VALIDATE:  91%|█████████▏| 117/128 [01:37<00:25,  2.30s/it]
cache-cache_patch:transformers-rt/inputs2 VALIDATE:  92%|█████████▏| 118/128 [01:37<00:33,  3.33s/it]
cache-cache_patch:transformers-rt/inputs_empty_cache EXPORT:  92%|█████████▏| 118/128 [01:37<00:33,  3.33s/it]
cache-cache_patch:transformers-rt/inputs_empty_cache EXPORT:  93%|█████████▎| 119/128 [01:38<00:23,  2.56s/it]
cache-cache_patch:transformers-rt/inputs_batch1 EXPORT:  93%|█████████▎| 119/128 [01:38<00:23,  2.56s/it]
cache-cache_patch:transformers-rt/inputs_batch1 EXPORT:  94%|█████████▍| 120/128 [01:39<00:16,  2.01s/it]
cache-cache_patch:transformers-oblivious/inputs EXPORT:  94%|█████████▍| 120/128 [01:39<00:16,  2.01s/it][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-oblivious/inputs VALIDATE:  94%|█████████▍| 120/128 [01:43<00:16,  2.01s/it]
cache-cache_patch:transformers-oblivious/inputs VALIDATE:  95%|█████████▍| 121/128 [01:43<00:19,  2.75s/it]
cache-cache_patch:transformers-oblivious/inputs2 EXPORT:  95%|█████████▍| 121/128 [01:43<00:19,  2.75s/it] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-oblivious/inputs2 VALIDATE:  95%|█████████▍| 121/128 [01:47<00:19,  2.75s/it]
cache-cache_patch:transformers-oblivious/inputs2 VALIDATE:  95%|█████████▌| 122/128 [01:48<00:19,  3.28s/it]
cache-cache_patch:transformers-oblivious/inputs_empty_cache EXPORT:  95%|█████████▌| 122/128 [01:48<00:19,  3.28s/it][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-oblivious/inputs_empty_cache VALIDATE:  95%|█████████▌| 122/128 [01:52<00:19,  3.28s/it]
cache-cache_patch:transformers-oblivious/inputs_empty_cache VALIDATE:  96%|█████████▌| 123/128 [01:52<00:18,  3.62s/it]
cache-cache_patch:transformers-oblivious/inputs_batch1 EXPORT:  96%|█████████▌| 123/128 [01:52<00:18,  3.62s/it]       [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ❌

cache-cache_patch:transformers-oblivious/inputs_batch1 EXPORT:  97%|█████████▋| 124/128 [01:57<00:16,  4.11s/it]
cache-cache_patch:transformers-oblivious-rt/inputs EXPORT:  97%|█████████▋| 124/128 [01:57<00:16,  4.11s/it]    [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-oblivious-rt/inputs VALIDATE:  97%|█████████▋| 124/128 [02:00<00:16,  4.11s/it]
cache-cache_patch:transformers-oblivious-rt/inputs VALIDATE:  98%|█████████▊| 125/128 [02:01<00:11,  3.87s/it]
cache-cache_patch:transformers-oblivious-rt/inputs2 EXPORT:  98%|█████████▊| 125/128 [02:01<00:11,  3.87s/it] [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-oblivious-rt/inputs2 VALIDATE:  98%|█████████▊| 125/128 [02:05<00:11,  3.87s/it]
cache-cache_patch:transformers-oblivious-rt/inputs2 VALIDATE:  98%|█████████▊| 126/128 [02:05<00:08,  4.09s/it]
cache-cache_patch:transformers-oblivious-rt/inputs_empty_cache EXPORT:  98%|█████████▊| 126/128 [02:05<00:08,  4.09s/it][torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: batch will not be used, since it shares the same shape constraints with another axis: batch.
  warnings.warn(
~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_dynamic_shapes.py:264: UserWarning: # The axis name: cache+seq will not be used, since it shares the same shape constraints with another axis: seq_length.
  warnings.warn(
Applied 39 of general pattern rewrite rules.

cache-cache_patch:transformers-oblivious-rt/inputs_empty_cache VALIDATE:  98%|█████████▊| 126/128 [02:09<00:08,  4.09s/it]
cache-cache_patch:transformers-oblivious-rt/inputs_empty_cache VALIDATE:  99%|█████████▉| 127/128 [02:10<00:04,  4.20s/it]
cache-cache_patch:transformers-oblivious-rt/inputs_batch1 EXPORT:  99%|█████████▉| 127/128 [02:10<00:04,  4.20s/it]       [torch.onnx] Run decomposition...
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ❌

cache-cache_patch:transformers-oblivious-rt/inputs_batch1 EXPORT: 100%|██████████| 128/128 [02:15<00:00,  4.61s/it]
cache-cache_patch:transformers-oblivious-rt/inputs_batch1 EXPORT: 100%|██████████| 128/128 [02:15<00:00,  1.06s/it]

Let’s save the results.

df = pandas.DataFrame(results)
df.to_excel("plot_export_tiny_llm_dim01_onnx.xlsx")
df
cache cache_patch oblivious rt export_with EXPORT ERR-EXPORT run_with WORKS ERR-RUN
0 0 0 0 0 inputs 0 8*s72 (133002505182080)is not tracked with pro... NaN NaN NaN
1 0 0 0 0 inputs2 0 8*s72 (133002504099056)is not tracked with pro... NaN NaN NaN
2 0 0 0 0 inputs_empty_cache 0 8*s72 (133002504421056)is not tracked with pro... NaN NaN NaN
3 0 0 0 0 inputs_batch1 0 8*s31 + 8*s70 (133002504843056)is not tracked ... NaN NaN NaN
4 0 0 0 1 inputs 0 8*s72 (133002503214656)is not tracked with pro... NaN NaN NaN
... ... ... ... ... ... ... ... ... ... ...
147 1 transformers 1 1 inputs_empty_cache 1 NaN inputs 1.0 NaN
148 1 transformers 1 1 inputs_empty_cache 1 NaN inputs2 1.0 NaN
149 1 transformers 1 1 inputs_empty_cache 1 NaN inputs_empty_cache 1.0 NaN
150 1 transformers 1 1 inputs_empty_cache 1 NaN inputs_batch1 1.0 NaN
151 1 transformers 1 1 inputs_batch1 0 Failed to convert the exported program to an O... NaN NaN NaN

152 rows × 10 columns



no_export = df[df.EXPORT == 0]
no_export.to_excel("plot_export_tiny_llm_dim01_onnx.no_export.xlsx")
no_export
cache cache_patch oblivious rt export_with EXPORT ERR-EXPORT run_with WORKS ERR-RUN
0 0 0 0 0 inputs 0 8*s72 (133002505182080)is not tracked with pro... NaN NaN NaN
1 0 0 0 0 inputs2 0 8*s72 (133002504099056)is not tracked with pro... NaN NaN NaN
2 0 0 0 0 inputs_empty_cache 0 8*s72 (133002504421056)is not tracked with pro... NaN NaN NaN
3 0 0 0 0 inputs_batch1 0 8*s31 + 8*s70 (133002504843056)is not tracked ... NaN NaN NaN
4 0 0 0 1 inputs 0 8*s72 (133002503214656)is not tracked with pro... NaN NaN NaN
5 0 0 0 1 inputs2 0 8*s72 (133002503573808)is not tracked with pro... NaN NaN NaN
6 0 0 0 1 inputs_empty_cache 0 8*s72 (133002503831776)is not tracked with pro... NaN NaN NaN
7 0 0 0 1 inputs_batch1 0 8*s31 + 8*s70 (133002502312912)is not tracked ... NaN NaN NaN
8 0 0 1 0 inputs 0 8*s72 (133002502832736)is not tracked with pro... NaN NaN NaN
9 0 0 1 0 inputs2 0 8*s72 (133002501060992)is not tracked with pro... NaN NaN NaN
10 0 0 1 0 inputs_empty_cache 0 8*s72 (133002501458112)is not tracked with pro... NaN NaN NaN
11 0 0 1 0 inputs_batch1 0 8*s72 (133002501857664)is not tracked with pro... NaN NaN NaN
12 0 0 1 1 inputs 0 8*s72 (133002500213168)is not tracked with pro... NaN NaN NaN
13 0 0 1 1 inputs2 0 8*s72 (133002500650816)is not tracked with pro... NaN NaN NaN
14 0 0 1 1 inputs_empty_cache 0 8*s72 (133002498991040)is not tracked with pro... NaN NaN NaN
15 0 0 1 1 inputs_batch1 0 8*s72 (133002499427936)is not tracked with pro... NaN NaN NaN
16 1 0 0 0 inputs 0 8*s72 (133002497766752)is not tracked with pro... NaN NaN NaN
17 1 0 0 0 inputs2 0 8*s72 (133002498267792)is not tracked with pro... NaN NaN NaN
18 1 0 0 0 inputs_empty_cache 0 8*s72 (133002498601040)is not tracked with pro... NaN NaN NaN
19 1 0 0 0 inputs_batch1 0 8*s31 + 8*s70 (133002497208304)is not tracked ... NaN NaN NaN
20 1 0 0 1 inputs 0 8*s72 (133002497663792)is not tracked with pro... NaN NaN NaN
21 1 0 0 1 inputs2 0 8*s72 (133002496044560)is not tracked with pro... NaN NaN NaN
22 1 0 0 1 inputs_empty_cache 0 8*s72 (133002496514752)is not tracked with pro... NaN NaN NaN
23 1 0 0 1 inputs_batch1 0 8*s31 + 8*s70 (133002495181744)is not tracked ... NaN NaN NaN
24 1 0 1 0 inputs 0 8*s72 (133002495537520)is not tracked with pro... NaN NaN NaN
25 1 0 1 0 inputs2 0 8*s72 (133002494032976)is not tracked with pro... NaN NaN NaN
26 1 0 1 0 inputs_empty_cache 0 8*s72 (133002494405200)is not tracked with pro... NaN NaN NaN
27 1 0 1 0 inputs_batch1 0 8*s72 (133002493012176)is not tracked with pro... NaN NaN NaN
28 1 0 1 1 inputs 0 8*s72 (133002491466848)is not tracked with pro... NaN NaN NaN
29 1 0 1 1 inputs2 0 8*s72 (133002492982096)is not tracked with pro... NaN NaN NaN
30 1 0 1 1 inputs_empty_cache 0 8*s72 (133002494897584)is not tracked with pro... NaN NaN NaN
31 1 0 1 1 inputs_batch1 0 8*s72 (133002495191200)is not tracked with pro... NaN NaN NaN
76 1 all 1 0 inputs_batch1 0 Failed to decompose the FX graph for ONNX comp... NaN NaN NaN
89 1 all 1 1 inputs_batch1 0 Failed to convert the exported program to an O... NaN NaN NaN
90 1 torch 0 0 inputs 0 Could not extract specialized integer from dat... NaN NaN NaN
91 1 torch 0 0 inputs2 0 Could not extract specialized integer from dat... NaN NaN NaN
92 1 torch 0 0 inputs_empty_cache 0 Could not extract specialized integer from dat... NaN NaN NaN
93 1 torch 0 0 inputs_batch1 0 Could not extract specialized integer from dat... NaN NaN NaN
94 1 torch 0 1 inputs 0 Could not extract specialized integer from dat... NaN NaN NaN
95 1 torch 0 1 inputs2 0 Could not extract specialized integer from dat... NaN NaN NaN
96 1 torch 0 1 inputs_empty_cache 0 Could not extract specialized integer from dat... NaN NaN NaN
97 1 torch 0 1 inputs_batch1 0 Could not extract specialized integer from dat... NaN NaN NaN
98 1 torch 1 0 inputs 0 Could not extract specialized integer from dat... NaN NaN NaN
99 1 torch 1 0 inputs2 0 Could not extract specialized integer from dat... NaN NaN NaN
100 1 torch 1 0 inputs_empty_cache 0 Could not extract specialized integer from dat... NaN NaN NaN
101 1 torch 1 0 inputs_batch1 0 Could not extract specialized integer from dat... NaN NaN NaN
102 1 torch 1 1 inputs 0 Could not extract specialized integer from dat... NaN NaN NaN
103 1 torch 1 1 inputs2 0 Could not extract specialized integer from dat... NaN NaN NaN
104 1 torch 1 1 inputs_empty_cache 0 Could not extract specialized integer from dat... NaN NaN NaN
105 1 torch 1 1 inputs_batch1 0 Could not extract specialized integer from dat... NaN NaN NaN
114 1 transformers 0 0 inputs_empty_cache 0 Found the following conflicts between user-spe... NaN NaN NaN
115 1 transformers 0 0 inputs_batch1 0 Found the following conflicts between user-spe... NaN NaN NaN
124 1 transformers 0 1 inputs_empty_cache 0 Found the following conflicts between user-spe... NaN NaN NaN
125 1 transformers 0 1 inputs_batch1 0 Found the following conflicts between user-spe... NaN NaN NaN
138 1 transformers 1 0 inputs_batch1 0 Failed to convert the exported program to an O... NaN NaN NaN
151 1 transformers 1 1 inputs_batch1 0 Failed to convert the exported program to an O... NaN NaN NaN


The validation failures.

invalid = df[(df.EXPORT == 1) & (df.WORKS == 0)].pivot(
    index=["cache", "cache_patch", "oblivious", "rt", "export_with"],
    columns=["run_with"],
    values=["WORKS", "ERR-RUN"],
)
invalid.to_excel("plot_export_tiny_llm_dim01_onnx.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 oblivious rt export_with
1 all 0 0 inputs_batch1 0.0 0.0 NaN 0.0 [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... NaN [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go...
inputs_empty_cache 0.0 0.0 0.0 NaN [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... NaN
1 inputs_batch1 0.0 0.0 NaN 0.0 [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... NaN [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go...
inputs_empty_cache 0.0 0.0 0.0 NaN [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Go... NaN 


success = df[(df.EXPORT == 1) & (df.WORKS == 1)].pivot(
    index=["cache", "cache_patch", "oblivious", "rt", "export_with"],
    columns=["run_with"],
    values=["WORKS"],
)
success.to_excel("plot_export_tiny_llm_dim01_onnx.success.xlsx")
success
WORKS
run_with inputs inputs2 inputs_batch1 inputs_empty_cache
cache cache_patch oblivious rt export_with
1 all 0 0 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache NaN NaN NaN 1.0
1 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
inputs_batch1 NaN NaN 1.0 NaN
inputs_empty_cache NaN NaN NaN 1.0
1 0 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
inputs_empty_cache 1.0 1.0 1.0 1.0
1 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
inputs_empty_cache 1.0 1.0 1.0 1.0
transformers 0 0 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
1 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
1 0 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
inputs_empty_cache 1.0 1.0 1.0 1.0
1 inputs 1.0 1.0 1.0 1.0
inputs2 1.0 1.0 1.0 1.0
inputs_empty_cache 1.0 1.0 1.0 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.onnx.export", "tomato")
plot export tiny llm dim01 onnx

Total running time of the script: (2 minutes 20.810 seconds)

Related examples

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

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

Export with dynamic dimensions in {0,1}

Export with dynamic dimensions in {0,1}

Export microsoft/phi-2

Export microsoft/phi-2

Gallery generated by Sphinx-Gallery