Note
Go to the end to download the full example code.
Test the export on untrained models¶
Checking the exporter on a whole model takes time as it is usually big but we can create a smaller version with the same architecture. Then fix export issues on such a small model is faster.
codellama/CodeLlama-7b-Python-hf¶
Let’s grab some information about this model. This reuses huggingface_hub API.
import copy
import pprint
import torch
from onnx_diagnostic import doc
from onnx_diagnostic.ext_test_case import unit_test_going
from onnx_diagnostic.helpers import string_type
from onnx_diagnostic.torch_models.hghub import (
get_untrained_model_with_inputs,
)
from onnx_diagnostic.torch_models.hghub.hub_api import (
get_model_info,
get_pretrained_config,
task_from_id,
)
from onnx_diagnostic.torch_export_patches import torch_export_patches
from onnx_diagnostic.torch_export_patches.patch_inputs import use_dyn_not_str
model_id = (
"HuggingFaceM4/tiny-random-idefics"
if unit_test_going()
else "codellama/CodeLlama-7b-Python-hf"
)
print(f"model_id={model_id!r}")
print("info", get_model_info(model_id))
model_id='codellama/CodeLlama-7b-Python-hf'
info ModelInfo(id='codellama/CodeLlama-7b-Python-hf', author='codellama', sha='d4178f5d2eead875e627ec487b23679266319b7f', created_at=datetime.datetime(2023, 8, 24, 16, 31, 28, tzinfo=datetime.timezone.utc), last_modified=datetime.datetime(2024, 4, 12, 14, 16, 26, tzinfo=datetime.timezone.utc), private=False, disabled=False, downloads=6711, downloads_all_time=None, gated=False, gguf=None, inference=None, inference_provider_mapping=None, likes=144, library_name='transformers', tags=['transformers', 'pytorch', 'safetensors', 'llama', 'text-generation', 'llama-2', 'code', 'arxiv:2308.12950', 'license:llama2', 'autotrain_compatible', 'text-generation-inference', 'endpoints_compatible', 'region:us'], pipeline_tag='text-generation', mask_token=None, card_data={'base_model': None, 'datasets': None, 'eval_results': None, 'language': ['code'], 'library_name': None, 'license': 'llama2', 'license_name': None, 'license_link': None, 'metrics': None, 'model_name': None, 'pipeline_tag': 'text-generation', 'tags': ['llama-2']}, widget_data=None, model_index=None, config={'architectures': ['LlamaForCausalLM'], 'model_type': 'llama', 'tokenizer_config': {'bos_token': {'__type': 'AddedToken', 'content': '<s>', 'lstrip': False, 'normalized': True, 'rstrip': False, 'single_word': False}, 'eos_token': {'__type': 'AddedToken', 'content': '</s>', 'lstrip': False, 'normalized': True, 'rstrip': False, 'single_word': False}, 'pad_token': None, 'unk_token': {'__type': 'AddedToken', 'content': '<unk>', 'lstrip': False, 'normalized': True, 'rstrip': False, 'single_word': False}}}, transformers_info=TransformersInfo(auto_model='AutoModelForCausalLM', custom_class=None, pipeline_tag='text-generation', processor='AutoTokenizer'), trending_score=None, siblings=[RepoSibling(rfilename='.gitattributes', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='LICENSE', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='README.md', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='USE_POLICY.md', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='config.json', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='generation_config.json', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='model-00001-of-00002.safetensors', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='model-00002-of-00002.safetensors', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='model.safetensors.index.json', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='pytorch_model-00001-of-00003.bin', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='pytorch_model-00002-of-00003.bin', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='pytorch_model-00003-of-00003.bin', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='pytorch_model.bin.index.json', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='special_tokens_map.json', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='tokenizer.json', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='tokenizer.model', size=None, blob_id=None, lfs=None), RepoSibling(rfilename='tokenizer_config.json', size=None, blob_id=None, lfs=None)], spaces=['bigcode/bigcode-models-leaderboard', 'Intel/low_bit_open_llm_leaderboard', 'BAAI/open_cn_llm_leaderboard', 'qiantong-xu/toolbench-leaderboard', 'gsaivinay/open_llm_leaderboard', 'EvanTHU/MotionLLM', 'GTBench/GTBench', 'Vikhrmodels/small-shlepa-lb', 'Vikhrmodels/DOoM-lb', 'kz-transformers/kaz-llm-lb', '21world/bigcode-models-leaderboard', 'felixz/open_llm_leaderboard', 'BAAI/open_flageval_vlm_leaderboard', 'HemaAM/GPT_train_on_LLaMa', 'OPTML-Group/UnlearnCanvas-Benchmark', 'atlasas/bigcode-models-leaderboard', 'whackthejacker/CodeTuneStudio', 'anantgupta129/LitGPT-Pythia-160M', 'BAAI/EmbodiedVerse', 'neubla/neubla-llm-evaluation-board', 'PrarthanaTS/tsai-gpt-from-scratch', 'MadhurGarg/TSAIGPTRedPajama', 'RaviNaik/ERA-SESSION22', 'theangkko/codellama-CodeLlama-7b-Python-hf', 'rodrigomasini/data_only_open_llm_leaderboard', 'Docfile/open_llm_leaderboard', 'Sijuade/GPTNEXTWORD', 'VDebugger/VDebugger-generalist-for-VQA', 'Temuzin64/code_helper', 'Agents-MCP-Hackathon/universal-api-translator', 'Noveumai/NovaEval', 'piyushgrover/MiniGPT_S22', 'supra-e-acc/Pythia-160M-text-generate', 'venkyyuvy/GPT_redpajama', 'mkthoma/GPT_From_Scratch', 'VarunSivamani/GPT-From-Scratch', 'sanjanatule/GPTNext', 'RashiAgarwal/TSAIGPTRedPajama', 'neuralorbs/DialogGen', 'Navyabhat/ERAV1-Session-22', 'GunaKoppula/ERA-Session-22', 'Vaish2705/ERA_S22', 'smothiki/open_llm_leaderboard', 'aemonge/codellama-CodeLlama-7b-Python-hf', 'sid92/codellama-CodeLlama-7b-Python-hf', 'poubellearman/codellama-CodeLlama-7b-Python-hf', 'IPF/codellama-CodeLlama-7b-Python-hf', 'Chris4K/codellama-CodeLlama-7b-Python-hf', 'CuriosityPdf/codellama-CodeLlama-7b-Python-hf', 'shreefhamed/codellama-CodeLlama-7b-Python-hf', 'markl11/codellama-CodeLlama-7b-Python-hf', '0x1668/open_llm_leaderboard', 'rpratl/codellama-CodeLlama-7b-Python-hf', 'pngwn/open_llm_leaderboard-check', 'asir0z/open_llm_leaderboard', 'LovelySweet/codellama-CodeLlama-7b-Python-hf', 'kbmlcoding/open_llm_leaderboard_free', 'ToletiSri/TSAI_S22', 'aichampions/open_llm_leaderboard', 'Adeco/open_llm_leaderboard', 'anirudh937/open_llm_leaderboard', 'smothiki/open_llm_leaderboard2', 'mjalg/IFEvalTR', 'lastsamuraii/LitGPT-Pythia-160M', 'Wazahat/Pyco', 'feryelb/python-coder', 'moshabann/Virtual_Teachers', 'ahmedsqrd/model_trace'], safetensors=SafeTensorsInfo(parameters={'BF16': 6738415616}, total=6738415616), security_repo_status=None, xet_enabled=None)
The configuration.
print("config", get_pretrained_config(model_id))
config LlamaConfig {
"architectures": [
"LlamaForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"bos_token_id": 1,
"dtype": "bfloat16",
"eos_token_id": 2,
"head_dim": 128,
"hidden_act": "silu",
"hidden_size": 4096,
"initializer_range": 0.02,
"intermediate_size": 11008,
"max_position_embeddings": 16384,
"mlp_bias": false,
"model_type": "llama",
"num_attention_heads": 32,
"num_hidden_layers": 32,
"num_key_value_heads": 32,
"pretraining_tp": 1,
"rms_norm_eps": 1e-05,
"rope_scaling": null,
"rope_theta": 1000000,
"tie_word_embeddings": false,
"transformers_version": "4.56.0.dev0",
"use_cache": true,
"vocab_size": 32000
}
The task determines the set of inputs which needs to be created for this input.
print("task", task_from_id(model_id))
task text-generation
Untrained model¶
The function get_untrained_model_with_inputs.
It loads the pretrained configuration, extracts the task associated
to the model and them creates random inputs and dynamic shapes
for torch.export.export().
[get_untrained_model_with_inputs] model_id='codellama/CodeLlama-7b-Python-hf'
[get_untrained_model_with_inputs] use preinstalled 'codellama/CodeLlama-7b-Python-hf'
[get_untrained_model_with_inputs] architectures=['LlamaForCausalLM']
[get_untrained_model_with_inputs] cls='LlamaConfig'
[get_untrained_model_with_inputs] task='text-generation'
[get_untrained_model_with_inputs] default config._attn_implementation=None
[get_untrained_model_with_inputs] use fct=<function get_inputs at 0x711526ee6700>
model size: 2667659264
number of weights: 666914816
fields: {'configuration', 'model_kwargs', 'dynamic_shapes', 'inputs2', 'n_weights', 'input_kwargs', 'inputs', 'model', 'size', 'task'}
Inputs
print("inputs:", string_type(data["inputs"], with_shape=True))
inputs: dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#2[T1s2x32x30x128,T1s2x32x30x128], value_cache=#2[T1s2x32x30x128,T1s2x32x30x128]))
Dynamic Shapes
print("dynamic shapes:", pprint.pformat(data["dynamic_shapes"]))
dynamic shapes: {'attention_mask': {0: Dim('batch', min=1, max=1024), 1: 'cache+seq'},
'input_ids': {0: Dim('batch', min=1, max=1024), 1: 'seq_length'},
'past_key_values': [[{0: Dim('batch', min=1, max=1024), 2: 'cache_length'},
{0: Dim('batch', min=1, max=1024), 2: 'cache_length'}],
[{0: Dim('batch', min=1, max=1024), 2: 'cache_length'},
{0: Dim('batch', min=1, max=1024), 2: 'cache_length'}]],
'position_ids': {0: Dim('batch', min=1, max=1024), 1: 'cache+seq'}}
Let’s check the model runs. We still needs to copy the inputs before using the models, the cache is usually modified inplace. Expected outputs can be used later to compute discrepancies.
inputs_copy = copy.deepcopy(data["inputs"])
model = data["model"]
expected_outputs = model(**inputs_copy)
print("outputs:", string_type(expected_outputs, with_shape=True))
outputs: CausalLMOutputWithPast(logits:T1s2x3x32000,past_key_values:DynamicCache(key_cache=#2[T1s2x32x33x128,T1s2x32x33x128], value_cache=#2[T1s2x32x33x128,T1s2x32x33x128]))
It works.
Export¶
The model uses transformers.cache_utils.DynamicCache.
It still requires patches to be exportable (control flow).
See onnx_diagnostic.torch_export_patches.torch_export_patches()
with torch_export_patches(patch_transformers=True) as f:
ep = torch.export.export(
model,
(),
kwargs=f(data["inputs"]),
dynamic_shapes=use_dyn_not_str(data["dynamic_shapes"]),
strict=False,
)
print(ep)
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 4096]", p_model_layers_0_self_attn_q_proj_weight: "f32[4096, 4096]", p_model_layers_0_self_attn_k_proj_weight: "f32[4096, 4096]", p_model_layers_0_self_attn_v_proj_weight: "f32[4096, 4096]", p_model_layers_0_self_attn_o_proj_weight: "f32[4096, 4096]", p_model_layers_0_mlp_gate_proj_weight: "f32[11008, 4096]", p_model_layers_0_mlp_up_proj_weight: "f32[11008, 4096]", p_model_layers_0_mlp_down_proj_weight: "f32[4096, 11008]", p_model_layers_0_input_layernorm_weight: "f32[4096]", p_model_layers_0_post_attention_layernorm_weight: "f32[4096]", p_model_layers_1_self_attn_q_proj_weight: "f32[4096, 4096]", p_model_layers_1_self_attn_k_proj_weight: "f32[4096, 4096]", p_model_layers_1_self_attn_v_proj_weight: "f32[4096, 4096]", p_model_layers_1_self_attn_o_proj_weight: "f32[4096, 4096]", p_model_layers_1_mlp_gate_proj_weight: "f32[11008, 4096]", p_model_layers_1_mlp_up_proj_weight: "f32[11008, 4096]", p_model_layers_1_mlp_down_proj_weight: "f32[4096, 11008]", p_model_layers_1_input_layernorm_weight: "f32[4096]", p_model_layers_1_post_attention_layernorm_weight: "f32[4096]", p_model_norm_weight: "f32[4096]", p_lm_head_weight: "f32[32000, 4096]", b_model_rotary_emb_inv_freq: "f32[64]", input_ids: "i64[s23, s70]", attention_mask: "i64[s23, s53]", position_ids: "i64[s23, s70]", past_key_values_key_cache_0: "f32[s23, 32, s31, 128]", past_key_values_key_cache_1: "f32[s23, 32, s31, 128]", past_key_values_value_cache_0: "f32[s23, 32, s11, 128]", past_key_values_value_cache_1: "f32[s23, 32, s54, 128]"):
#
sym_size_int_16: "Sym(s70)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_21: "Sym(s23)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 0)
sym_size_int_22: "Sym(s31)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# No stacktrace found for following nodes
empty: "f32[s23, 32, 0, 128]" = torch.ops.aten.empty.memory_format([sym_size_int_21, 32, 0, 128], dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
empty_1: "f32[s23, 32, 0, 128]" = torch.ops.aten.empty.memory_format([sym_size_int_21, 32, 0, 128], dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
cat: "f32[s23, 32, s31, 128]" = torch.ops.aten.cat.default([empty, past_key_values_key_cache_0], -2); empty = past_key_values_key_cache_0 = None
cat_1: "f32[s23, 32, s11, 128]" = torch.ops.aten.cat.default([empty_1, past_key_values_value_cache_0], -2); empty_1 = past_key_values_value_cache_0 = None
empty_2: "f32[s23, 32, 0, 128]" = torch.ops.aten.empty.memory_format([sym_size_int_21, 32, 0, 128], dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
empty_3: "f32[s23, 32, 0, 128]" = torch.ops.aten.empty.memory_format([sym_size_int_21, 32, 0, 128], dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
cat_2: "f32[s23, 32, s31, 128]" = torch.ops.aten.cat.default([empty_2, past_key_values_key_cache_1], -2); empty_2 = past_key_values_key_cache_1 = None
cat_3: "f32[s23, 32, s54, 128]" = torch.ops.aten.cat.default([empty_3, past_key_values_value_cache_1], -2); empty_3 = past_key_values_value_cache_1 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:192 in forward, code: return F.embedding(
embedding: "f32[s23, s70, 4096]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:376 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s31 + s70)" = sym_size_int_22 + sym_size_int_16
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:375 in forward, code: cache_position: torch.Tensor = torch.arange(
arange: "i64[s70]" = torch.ops.aten.arange.start(sym_size_int_22, add, device = device(type='cpu'), pin_memory = False); sym_size_int_22 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:382 in forward, code: causal_mask = create_causal_mask(
_assert_tensor_metadata_default = torch.ops.aten._assert_tensor_metadata.default(attention_mask, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default = None
to: "b8[s23, s53]" = torch.ops.aten.to.device(attention_mask, device(type='cpu'), torch.bool); attention_mask = None
arange_1: "i64[s31 + s70]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
add_: "i64[s31 + s70]" = torch.ops.aten.add_.Tensor(arange_1, 0); arange_1 = None
arange_2: "i64[s23]" = torch.ops.aten.arange.default(sym_size_int_21, device = device(type='cpu'), pin_memory = False)
arange_3: "i64[1]" = torch.ops.aten.arange.default(1, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s23, 1, 1, 1]" = torch.ops.aten.reshape.default(arange_2, [-1, 1, 1, 1]); arange_2 = None
reshape_1: "i64[1, 1, 1, 1]" = torch.ops.aten.reshape.default(arange_3, [1, -1, 1, 1]); arange_3 = None
reshape_2: "i64[1, 1, s70, 1]" = torch.ops.aten.reshape.default(arange, [1, 1, -1, 1]); arange = None
reshape_3: "i64[1, 1, 1, s31 + s70]" = torch.ops.aten.reshape.default(add_, [1, 1, 1, -1]); add_ = None
expand: "i64[s23, 1, s70, s31 + s70]" = torch.ops.aten.expand.default(reshape, [sym_size_int_21, 1, sym_size_int_16, add]); reshape = None
expand_1: "i64[s23, 1, s70, s31 + s70]" = torch.ops.aten.expand.default(reshape_1, [sym_size_int_21, 1, sym_size_int_16, add]); reshape_1 = expand_1 = None
expand_2: "i64[s23, 1, s70, s31 + s70]" = torch.ops.aten.expand.default(reshape_2, [sym_size_int_21, 1, sym_size_int_16, add]); reshape_2 = None
expand_3: "i64[s23, 1, s70, s31 + s70]" = torch.ops.aten.expand.default(reshape_3, [sym_size_int_21, 1, sym_size_int_16, add]); reshape_3 = None
new_ones: "b8[]" = torch.ops.aten.new_ones.default(expand_2, [], dtype = torch.bool, pin_memory = False)
le: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.le.Tensor(expand_3, expand_2); expand_2 = None
_assert_tensor_metadata_default_1 = torch.ops.aten._assert_tensor_metadata.default(le, dtype = torch.bool, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_1 = None
to_1: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.to.dtype_layout(le, dtype = torch.bool, layout = torch.strided, device = device(type='cpu')); le = None
and_1: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.__and__.Tensor(new_ones, to_1); new_ones = to_1 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/_dynamo/_trace_wrapped_higher_order_op.py:99 in forward, code: return torch.ops.aten.index(x, indices)
index: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.index.Tensor(to, [expand, expand_3]); to = expand = expand_3 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:382 in forward, code: causal_mask = create_causal_mask(
_assert_tensor_metadata_default_2 = torch.ops.aten._assert_tensor_metadata.default(index, dtype = torch.bool, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_2 = None
to_2: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.to.dtype_layout(index, dtype = torch.bool, layout = torch.strided, device = device(type='cpu')); index = None
and_2: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.__and__.Tensor(and_1, to_2); and_1 = to_2 = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1040 in forward, code: self.inv_freq[None, :, None]
unsqueeze: "f32[1, 64]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
unsqueeze_1: "f32[1, 64, 1]" = torch.ops.aten.unsqueeze.default(unsqueeze, 2); unsqueeze = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1041 in forward, code: .float()
_assert_tensor_metadata_default_3 = torch.ops.aten._assert_tensor_metadata.default(unsqueeze_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_3 = None
to_3: "f32[1, 64, 1]" = torch.ops.aten.to.dtype(unsqueeze_1, torch.float32); unsqueeze_1 = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1042 in forward, code: .expand(position_ids.shape[0], -1, 1)
expand_4: "f32[s23, 64, 1]" = torch.ops.aten.expand.default(to_3, [sym_size_int_21, -1, 1]); to_3 = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1043 in forward, code: .to(x.device)
_assert_tensor_metadata_default_4 = torch.ops.aten._assert_tensor_metadata.default(expand_4, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_4 = None
to_4: "f32[s23, 64, 1]" = torch.ops.aten.to.dtype_layout(expand_4, dtype = torch.float32, layout = torch.strided, device = device(type='cpu')); expand_4 = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1045 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
unsqueeze_2: "i64[s23, 1, s70]" = torch.ops.aten.unsqueeze.default(position_ids, 1); position_ids = None
slice_1: "i64[s23, 1, s70]" = torch.ops.aten.slice.Tensor(unsqueeze_2, 2, 0, 9223372036854775807); unsqueeze_2 = None
_assert_tensor_metadata_default_5 = torch.ops.aten._assert_tensor_metadata.default(slice_1, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_5 = None
to_5: "f32[s23, 1, s70]" = torch.ops.aten.to.dtype(slice_1, torch.float32); slice_1 = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_autocast = torch.ops.higher_order.wrap_with_autocast('cpu', torch.bfloat16, False, False, submod_3, to_4, to_5); submod_3 = to_4 = to_5 = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1055 in forward, code: cos = emb.cos() * self.attention_scaling
mul: "f32[s23, s70, 128]" = wrap_with_autocast[0]
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1056 in forward, code: sin = emb.sin() * self.attention_scaling
mul_1: "f32[s23, s70, 128]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1058 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
_assert_tensor_metadata_default_8 = torch.ops.aten._assert_tensor_metadata.default(mul, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_8 = None
to_8: "f32[s23, s70, 128]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
_assert_tensor_metadata_default_9 = torch.ops.aten._assert_tensor_metadata.default(mul_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_9 = None
to_9: "f32[s23, s70, 128]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:64 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_10 = torch.ops.aten._assert_tensor_metadata.default(embedding, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_10 = None
to_10: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:65 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s23, s70, 4096]" = torch.ops.aten.pow.Tensor_Scalar(to_10, 2)
mean: "f32[s23, s70, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:66 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s23, s70, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s23, s70, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(to_10, rsqrt); rsqrt = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:67 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_11 = torch.ops.aten._assert_tensor_metadata.default(mul_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_11 = None
to_11: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_11); p_model_layers_0_input_layernorm_weight = to_11 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_q_proj_weight); p_model_layers_0_self_attn_q_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:236 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s23, s70, 32, 128]" = torch.ops.aten.view.default(linear, [sym_size_int_21, sym_size_int_16, -1, 128]); linear = None
transpose_1: "f32[s23, 32, s70, 128]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_k_proj_weight); p_model_layers_0_self_attn_k_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:237 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s23, s70, 32, 128]" = torch.ops.aten.view.default(linear_1, [sym_size_int_21, sym_size_int_16, -1, 128]); linear_1 = None
transpose_2: "f32[s23, 32, s70, 128]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_v_proj_weight); mul_3 = p_model_layers_0_self_attn_v_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:238 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s23, s70, 32, 128]" = torch.ops.aten.view.default(linear_2, [sym_size_int_21, sym_size_int_16, -1, 128]); linear_2 = None
transpose_3: "f32[s23, 32, s70, 128]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:241 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_3: "f32[s23, 1, s70, 128]" = torch.ops.aten.unsqueeze.default(to_8, 1)
unsqueeze_4: "f32[s23, 1, s70, 128]" = torch.ops.aten.unsqueeze.default(to_9, 1)
mul_4: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_3)
slice_2: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 64)
slice_3: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 64, 9223372036854775807); transpose_1 = None
neg: "f32[s23, 32, s70, 64]" = torch.ops.aten.neg.default(slice_3); slice_3 = None
cat_5: "f32[s23, 32, s70, 128]" = torch.ops.aten.cat.default([neg, slice_2], -1); neg = slice_2 = None
mul_5: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(cat_5, unsqueeze_4); cat_5 = None
add_4: "f32[s23, 32, s70, 128]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_3); unsqueeze_3 = None
slice_4: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 64)
slice_5: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 64, 9223372036854775807); transpose_2 = None
neg_1: "f32[s23, 32, s70, 64]" = torch.ops.aten.neg.default(slice_5); slice_5 = None
cat_6: "f32[s23, 32, s70, 128]" = torch.ops.aten.cat.default([neg_1, slice_4], -1); neg_1 = slice_4 = None
mul_7: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(cat_6, unsqueeze_4); cat_6 = unsqueeze_4 = None
add_5: "f32[s23, 32, s70, 128]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:246 in forward, code: key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_7: "f32[s23, 32, s31 + s70, 128]" = torch.ops.aten.cat.default([cat, add_5], -2); cat = add_5 = None
cat_8: "f32[s23, 32, s11 + s70, 128]" = torch.ops.aten.cat.default([cat_1, transpose_3], -2); cat_1 = transpose_3 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:252 in forward, code: attn_output, attn_weights = attention_interface(
slice_6: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.slice.Tensor(and_2, 3, None, add)
scaled_dot_product_attention: "f32[s23, 32, s70, 128]" = torch.ops.aten.scaled_dot_product_attention.default(add_4, cat_7, cat_8, slice_6, scale = 0.08838834764831845); add_4 = slice_6 = None
transpose_4: "f32[s23, s70, 32, 128]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:263 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_4: "f32[s23, s70, 4096]" = torch.ops.aten.reshape.default(transpose_4, [sym_size_int_21, sym_size_int_16, -1]); transpose_4 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(reshape_4, p_model_layers_0_self_attn_o_proj_weight); reshape_4 = p_model_layers_0_self_attn_o_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:304 in forward, code: hidden_states = residual + hidden_states
add_6: "f32[s23, s70, 4096]" = torch.ops.aten.add.Tensor(to_10, linear_3); to_10 = linear_3 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:64 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_12 = torch.ops.aten._assert_tensor_metadata.default(add_6, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_12 = None
to_12: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(add_6, torch.float32); add_6 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:65 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s23, s70, 4096]" = torch.ops.aten.pow.Tensor_Scalar(to_12, 2)
mean_1: "f32[s23, s70, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:66 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_7: "f32[s23, s70, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s23, s70, 1]" = torch.ops.aten.rsqrt.default(add_7); add_7 = None
mul_16: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(to_12, rsqrt_1); rsqrt_1 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:67 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_13 = torch.ops.aten._assert_tensor_metadata.default(mul_16, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_13 = None
to_13: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(mul_16, torch.float32); mul_16 = None
mul_17: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_13); p_model_layers_0_post_attention_layernorm_weight = to_13 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s23, s70, 11008]" = torch.ops.aten.linear.default(mul_17, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:473 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s23, s70, 11008]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s23, s70, 11008]" = torch.ops.aten.linear.default(mul_17, p_model_layers_0_mlp_up_proj_weight); mul_17 = p_model_layers_0_mlp_up_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:155 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_18: "f32[s23, s70, 11008]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_18, p_model_layers_0_mlp_down_proj_weight); mul_18 = p_model_layers_0_mlp_down_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:310 in forward, code: hidden_states = residual + hidden_states
add_8: "f32[s23, s70, 4096]" = torch.ops.aten.add.Tensor(to_12, linear_6); to_12 = linear_6 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:64 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_14 = torch.ops.aten._assert_tensor_metadata.default(add_8, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_14 = None
to_14: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(add_8, torch.float32); add_8 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:65 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s23, s70, 4096]" = torch.ops.aten.pow.Tensor_Scalar(to_14, 2)
mean_2: "f32[s23, s70, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:66 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_9: "f32[s23, s70, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s23, s70, 1]" = torch.ops.aten.rsqrt.default(add_9); add_9 = None
mul_19: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(to_14, rsqrt_2); rsqrt_2 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:67 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_15 = torch.ops.aten._assert_tensor_metadata.default(mul_19, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_15 = None
to_15: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(mul_19, torch.float32); mul_19 = None
mul_20: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(p_model_layers_1_input_layernorm_weight, to_15); p_model_layers_1_input_layernorm_weight = to_15 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_20, p_model_layers_1_self_attn_q_proj_weight); p_model_layers_1_self_attn_q_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:236 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_3: "f32[s23, s70, 32, 128]" = torch.ops.aten.view.default(linear_7, [sym_size_int_21, sym_size_int_16, -1, 128]); linear_7 = None
transpose_5: "f32[s23, 32, s70, 128]" = torch.ops.aten.transpose.int(view_3, 1, 2); view_3 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_8: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_20, p_model_layers_1_self_attn_k_proj_weight); p_model_layers_1_self_attn_k_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:237 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_4: "f32[s23, s70, 32, 128]" = torch.ops.aten.view.default(linear_8, [sym_size_int_21, sym_size_int_16, -1, 128]); linear_8 = None
transpose_6: "f32[s23, 32, s70, 128]" = torch.ops.aten.transpose.int(view_4, 1, 2); view_4 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_9: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_20, p_model_layers_1_self_attn_v_proj_weight); mul_20 = p_model_layers_1_self_attn_v_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:238 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_5: "f32[s23, s70, 32, 128]" = torch.ops.aten.view.default(linear_9, [sym_size_int_21, sym_size_int_16, -1, 128]); linear_9 = None
transpose_7: "f32[s23, 32, s70, 128]" = torch.ops.aten.transpose.int(view_5, 1, 2); view_5 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:241 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_5: "f32[s23, 1, s70, 128]" = torch.ops.aten.unsqueeze.default(to_8, 1); to_8 = None
unsqueeze_6: "f32[s23, 1, s70, 128]" = torch.ops.aten.unsqueeze.default(to_9, 1); to_9 = None
mul_21: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(transpose_5, unsqueeze_5)
slice_7: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_5, 3, 0, 64)
slice_8: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_5, 3, 64, 9223372036854775807); transpose_5 = None
neg_2: "f32[s23, 32, s70, 64]" = torch.ops.aten.neg.default(slice_8); slice_8 = None
cat_9: "f32[s23, 32, s70, 128]" = torch.ops.aten.cat.default([neg_2, slice_7], -1); neg_2 = slice_7 = None
mul_22: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(cat_9, unsqueeze_6); cat_9 = None
add_10: "f32[s23, 32, s70, 128]" = torch.ops.aten.add.Tensor(mul_21, mul_22); mul_21 = mul_22 = None
mul_23: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(transpose_6, unsqueeze_5); unsqueeze_5 = None
slice_9: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_6, 3, 0, 64)
slice_10: "f32[s23, 32, s70, 64]" = torch.ops.aten.slice.Tensor(transpose_6, 3, 64, 9223372036854775807); transpose_6 = None
neg_3: "f32[s23, 32, s70, 64]" = torch.ops.aten.neg.default(slice_10); slice_10 = None
cat_10: "f32[s23, 32, s70, 128]" = torch.ops.aten.cat.default([neg_3, slice_9], -1); neg_3 = slice_9 = None
mul_24: "f32[s23, 32, s70, 128]" = torch.ops.aten.mul.Tensor(cat_10, unsqueeze_6); cat_10 = unsqueeze_6 = None
add_11: "f32[s23, 32, s70, 128]" = torch.ops.aten.add.Tensor(mul_23, mul_24); mul_23 = mul_24 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:246 in forward, code: key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_11: "f32[s23, 32, s31 + s70, 128]" = torch.ops.aten.cat.default([cat_2, add_11], -2); cat_2 = add_11 = None
cat_12: "f32[s23, 32, s54 + s70, 128]" = torch.ops.aten.cat.default([cat_3, transpose_7], -2); cat_3 = transpose_7 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:252 in forward, code: attn_output, attn_weights = attention_interface(
slice_11: "b8[s23, 1, s70, s31 + s70]" = torch.ops.aten.slice.Tensor(and_2, 3, None, add); and_2 = add = None
scaled_dot_product_attention_1: "f32[s23, 32, s70, 128]" = torch.ops.aten.scaled_dot_product_attention.default(add_10, cat_11, cat_12, slice_11, scale = 0.08838834764831845); add_10 = slice_11 = None
transpose_8: "f32[s23, s70, 32, 128]" = torch.ops.aten.transpose.int(scaled_dot_product_attention_1, 1, 2); scaled_dot_product_attention_1 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:263 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_5: "f32[s23, s70, 4096]" = torch.ops.aten.reshape.default(transpose_8, [sym_size_int_21, sym_size_int_16, -1]); transpose_8 = sym_size_int_21 = sym_size_int_16 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_10: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(reshape_5, p_model_layers_1_self_attn_o_proj_weight); reshape_5 = p_model_layers_1_self_attn_o_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:304 in forward, code: hidden_states = residual + hidden_states
add_12: "f32[s23, s70, 4096]" = torch.ops.aten.add.Tensor(to_14, linear_10); to_14 = linear_10 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:64 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_16 = torch.ops.aten._assert_tensor_metadata.default(add_12, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_16 = None
to_16: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(add_12, torch.float32); add_12 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:65 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_4: "f32[s23, s70, 4096]" = torch.ops.aten.pow.Tensor_Scalar(to_16, 2)
mean_3: "f32[s23, s70, 1]" = torch.ops.aten.mean.dim(pow_4, [-1], True); pow_4 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:66 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_13: "f32[s23, s70, 1]" = torch.ops.aten.add.Tensor(mean_3, 1e-05); mean_3 = None
rsqrt_3: "f32[s23, s70, 1]" = torch.ops.aten.rsqrt.default(add_13); add_13 = None
mul_25: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(to_16, rsqrt_3); rsqrt_3 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:67 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_17 = torch.ops.aten._assert_tensor_metadata.default(mul_25, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_17 = None
to_17: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(mul_25, torch.float32); mul_25 = None
mul_26: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(p_model_layers_1_post_attention_layernorm_weight, to_17); p_model_layers_1_post_attention_layernorm_weight = to_17 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_11: "f32[s23, s70, 11008]" = torch.ops.aten.linear.default(mul_26, p_model_layers_1_mlp_gate_proj_weight); p_model_layers_1_mlp_gate_proj_weight = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:473 in forward, code: return F.silu(input, inplace=self.inplace)
silu_1: "f32[s23, s70, 11008]" = torch.ops.aten.silu.default(linear_11); linear_11 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_12: "f32[s23, s70, 11008]" = torch.ops.aten.linear.default(mul_26, p_model_layers_1_mlp_up_proj_weight); mul_26 = p_model_layers_1_mlp_up_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:155 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_27: "f32[s23, s70, 11008]" = torch.ops.aten.mul.Tensor(silu_1, linear_12); silu_1 = linear_12 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_13: "f32[s23, s70, 4096]" = torch.ops.aten.linear.default(mul_27, p_model_layers_1_mlp_down_proj_weight); mul_27 = p_model_layers_1_mlp_down_proj_weight = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:310 in forward, code: hidden_states = residual + hidden_states
add_14: "f32[s23, s70, 4096]" = torch.ops.aten.add.Tensor(to_16, linear_13); to_16 = linear_13 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:64 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_18 = torch.ops.aten._assert_tensor_metadata.default(add_14, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_18 = None
to_18: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(add_14, torch.float32); add_14 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:65 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_5: "f32[s23, s70, 4096]" = torch.ops.aten.pow.Tensor_Scalar(to_18, 2)
mean_4: "f32[s23, s70, 1]" = torch.ops.aten.mean.dim(pow_5, [-1], True); pow_5 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:66 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_15: "f32[s23, s70, 1]" = torch.ops.aten.add.Tensor(mean_4, 1e-05); mean_4 = None
rsqrt_4: "f32[s23, s70, 1]" = torch.ops.aten.rsqrt.default(add_15); add_15 = None
mul_28: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(to_18, rsqrt_4); to_18 = rsqrt_4 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:67 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_19 = torch.ops.aten._assert_tensor_metadata.default(mul_28, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_19 = None
to_19: "f32[s23, s70, 4096]" = torch.ops.aten.to.dtype(mul_28, torch.float32); mul_28 = None
mul_29: "f32[s23, s70, 4096]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_19); p_model_norm_weight = to_19 = None
# File: ~/github/transformers/src/transformers/models/llama/modeling_llama.py:473 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_12: "f32[s23, s70, 4096]" = torch.ops.aten.slice.Tensor(mul_29, 1, 0, 9223372036854775807); mul_29 = None
# File: ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:134 in forward, code: return F.linear(input, self.weight, self.bias)
linear_14: "f32[s23, s70, 32000]" = torch.ops.aten.linear.default(slice_12, p_lm_head_weight); slice_12 = p_lm_head_weight = None
return (linear_14, cat_7, cat_11, cat_8, cat_12)
class submod_1(torch.nn.Module):
def forward(self, to_4: "f32[s23, 64, 1]", to_5: "f32[s23, 1, s70]"):
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1053 in forward, code: freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
_assert_tensor_metadata_default_6 = torch.ops.aten._assert_tensor_metadata.default(to_4, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_6 = None
to_6: "f32[s23, 64, 1]" = torch.ops.aten.to.dtype(to_4, torch.float32); to_4 = None
_assert_tensor_metadata_default_7 = torch.ops.aten._assert_tensor_metadata.default(to_5, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_7 = None
to_7: "f32[s23, 1, s70]" = torch.ops.aten.to.dtype(to_5, torch.float32); to_5 = None
matmul: "f32[s23, 64, s70]" = torch.ops.aten.matmul.default(to_6, to_7); to_6 = to_7 = None
transpose: "f32[s23, s70, 64]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1054 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat_4: "f32[s23, s70, 128]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1055 in forward, code: cos = emb.cos() * self.attention_scaling
cos: "f32[s23, s70, 128]" = torch.ops.aten.cos.default(cat_4)
mul: "f32[s23, s70, 128]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: ~/github/onnx-diagnostic/onnx_diagnostic/torch_export_patches/patches/patch_transformers.py:1056 in forward, code: sin = emb.sin() * self.attention_scaling
sin: "f32[s23, s70, 128]" = torch.ops.aten.sin.default(cat_4); cat_4 = None
mul_1: "f32[s23, s70, 128]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
return (mul, mul_1)
Graph signature:
# inputs
p_model_embed_tokens_weight: PARAMETER target='model.embed_tokens.weight'
p_model_layers_0_self_attn_q_proj_weight: PARAMETER target='model.layers.0.self_attn.q_proj.weight'
p_model_layers_0_self_attn_k_proj_weight: PARAMETER target='model.layers.0.self_attn.k_proj.weight'
p_model_layers_0_self_attn_v_proj_weight: PARAMETER target='model.layers.0.self_attn.v_proj.weight'
p_model_layers_0_self_attn_o_proj_weight: PARAMETER target='model.layers.0.self_attn.o_proj.weight'
p_model_layers_0_mlp_gate_proj_weight: PARAMETER target='model.layers.0.mlp.gate_proj.weight'
p_model_layers_0_mlp_up_proj_weight: PARAMETER target='model.layers.0.mlp.up_proj.weight'
p_model_layers_0_mlp_down_proj_weight: PARAMETER target='model.layers.0.mlp.down_proj.weight'
p_model_layers_0_input_layernorm_weight: PARAMETER target='model.layers.0.input_layernorm.weight'
p_model_layers_0_post_attention_layernorm_weight: PARAMETER target='model.layers.0.post_attention_layernorm.weight'
p_model_layers_1_self_attn_q_proj_weight: PARAMETER target='model.layers.1.self_attn.q_proj.weight'
p_model_layers_1_self_attn_k_proj_weight: PARAMETER target='model.layers.1.self_attn.k_proj.weight'
p_model_layers_1_self_attn_v_proj_weight: PARAMETER target='model.layers.1.self_attn.v_proj.weight'
p_model_layers_1_self_attn_o_proj_weight: PARAMETER target='model.layers.1.self_attn.o_proj.weight'
p_model_layers_1_mlp_gate_proj_weight: PARAMETER target='model.layers.1.mlp.gate_proj.weight'
p_model_layers_1_mlp_up_proj_weight: PARAMETER target='model.layers.1.mlp.up_proj.weight'
p_model_layers_1_mlp_down_proj_weight: PARAMETER target='model.layers.1.mlp.down_proj.weight'
p_model_layers_1_input_layernorm_weight: PARAMETER target='model.layers.1.input_layernorm.weight'
p_model_layers_1_post_attention_layernorm_weight: PARAMETER target='model.layers.1.post_attention_layernorm.weight'
p_model_norm_weight: PARAMETER target='model.norm.weight'
p_lm_head_weight: PARAMETER target='lm_head.weight'
b_model_rotary_emb_inv_freq: BUFFER target='model.rotary_emb.inv_freq' persistent=False
input_ids: USER_INPUT
attention_mask: USER_INPUT
position_ids: USER_INPUT
past_key_values_key_cache_0: USER_INPUT
past_key_values_key_cache_1: USER_INPUT
past_key_values_value_cache_0: USER_INPUT
past_key_values_value_cache_1: USER_INPUT
# outputs
linear_14: USER_OUTPUT
cat_7: USER_OUTPUT
cat_11: USER_OUTPUT
cat_8: USER_OUTPUT
cat_12: USER_OUTPUT
Range constraints: {s23: VR[1, 1024], s70: VR[2, int_oo], s53: VR[4, int_oo], s31: VR[2, int_oo], s11: VR[2, int_oo], s54: VR[2, int_oo]}
doc.plot_legend(
"untrained\ncodellama/\nCodeLlama-7b-Python-hf", "torch.export.export", "tomato"
)
Total running time of the script: (0 minutes 7.231 seconds)
Related examples
Steel method forward to guess inputs and dynamic shapes (with Tiny-LLM)
Steel method forward to guess inputs and dynamic shapes (with Tiny-LLM)
Intermediate results with (ONNX) ReferenceEvaluator
Intermediate results with (ONNX) ReferenceEvaluator