Note
Go to the end to download the full example code.
Export Tiny-LLM with patches¶
Many models from transformers cannot be converted because the implementation uses cache classes. Let’s see how to get around that. We focus on the model arnir0/Tiny-LLM. To avoid downloading any weights, we write a function creating a random model based on the same architecture. This continues example Steel method forward to guess the dynamic shapes (with Tiny-LLM).
Errors¶
They depend on transformers version.
transformers>=4.40,<4.50 cannot serialize DynamicCache and cannot
map dynamic shapes to instances of DynamicCache. The following errors
would appear:
torch._dynamo.exc.UserError: Cannot associate shape
[[{0: <class '....batch'>, 2: <class '....cache_length'>}],
[{0: <class '....batch'>, 2: <class '....cache_length'>}]]
specified at `dynamic_shapes['past_key_values']`
to non-tensor type <class 'transformers.cache_utils.DynamicCache'>
at `inputs['past_key_values']` (expected None)
For more information about this error,
see: https://pytorch.org/docs/main/generated/exportdb/index.html#dynamic-shapes-validation
With transformers==4.50, it shows the following:
torch._dynamo.exc.UserError: Constraints violated (batch)!
For more information, run with TORCH_LOGS="+dynamic".
- Not all values of batch = L['args'][1]['input_ids'].size()[0]
in the specified range batch <= 1024 are valid
because batch was inferred to be a constant (2).
- Not all values of batch = L['args'][1]['attention_mask'].size()[0]
in the specified range batch <= 1024 are valid
because batch was inferred to be a constant (2).
- Not all values of batch = L['args'][1]['past_key_values']['key_cache'][0].size()[0]
in the specified range batch <= 1024 are valid
because batch was inferred to be a constant (2).
- Not all values of batch = L['args'][1]['past_key_values']['value_cache'][0].size()[0]
in the specified range batch <= 1024 are valid
because batch was inferred to be a constant (2).
Suggested fixes:
batch = 2
However, this package implements a patch mechanism with replaces the part causing these issues.
Note
restart after an export failure
If the export fails, it is better to start executing again, or restart the kernel if you are in the notebook. The export may leave torch in one unstable state.
import copy
import pprint
import torch
import transformers
from onnx_diagnostic import doc
from onnx_diagnostic.helpers.cache_helper import is_cache_dynamic_registered
from onnx_diagnostic.helpers import string_type
from onnx_diagnostic.torch_export_patches import bypass_export_some_errors
from onnx_diagnostic.torch_models.llms import get_tiny_llm
experiment = get_tiny_llm()
untrained_model, inputs, dynamic_shapes = (
experiment["model"],
experiment["inputs"],
experiment["dynamic_shapes"],
)
cloned_inputs = copy.deepcopy(inputs)
Let’s show this inputs, this was inferred in example Steel method forward to guess the dynamic shapes (with Tiny-LLM).
print(string_type(inputs, with_shape=True))
dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#1[T1s2x1x30x96], value_cache=#1[T1s2x1x30x96]))
And the dynamic shapes
{'attention_mask': {0: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.batch'>,
1: _DimHint(type=<_DimHintType.DYNAMIC: 3>)},
'input_ids': {0: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.batch'>,
1: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.seq_length'>},
'past_key_values': [[{0: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.batch'>,
2: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.cache_length'>}],
[{0: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.batch'>,
2: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.cache_length'>}]],
'position_ids': {0: <class 'onnx_diagnostic.torch_models.untrained.llm_tiny_llm.batch'>,
1: _DimHint(type=<_DimHintType.DYNAMIC: 3>)}}
Before exporting, we check transformers.cache_utils.DynamicCache
can serialized and deserialized otherwise torch.export.export()
fails.
print("-- DynamicCache registered: ", is_cache_dynamic_registered())
-- DynamicCache registered: True
If they are not registered, function func:onnx_diagnostic.torch_export_patches.bypass_export_some_errors should take care of it. Then we export.
with bypass_export_some_errors(patch_transformers=True, verbose=10) as modificator:
assert is_cache_dynamic_registered() # it must be true here
ep = torch.export.export(
untrained_model,
(),
kwargs=modificator(cloned_inputs),
dynamic_shapes=dynamic_shapes,
strict=False, # mandatory for torch==2.6
)
print("It worked:")
print(ep)
[bypass_export_some_errors] replace torch.jit.isinstance, torch._dynamo.mark_static_address
[_register_cache_serialization] register MambaCache
[_register_cache_serialization] DynamicCache is unregistered first.
[_register_cache_serialization] register DynamicCache
[bypass_export_some_errors] patch sympy
[bypass_export_some_errors] patch pytorch
[bypass_export_some_errors] modifies shape constraints
[patch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_DynamicCache: reorder_cache, update, crop, from_batch_splits, get_seq_length
[patch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_GenerationMixin: _cache_dependant_input_preparation, _cache_dependant_input_preparation_exporting, prepare_inputs_for_generation
[bypass_export_some_errors] done patching
/home/xadupre/vv/this312/lib/python3.12/site-packages/torch/backends/mkldnn/__init__.py:78: UserWarning: TF32 acceleration on top of oneDNN is available for Intel GPUs. The current Torch version does not have Intel GPU Support. (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:148.)
torch._C._set_onednn_allow_tf32(_allow_tf32)
It worked:
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 192]", p_model_layers_0_self_attn_q_proj_weight: "f32[192, 192]", p_model_layers_0_self_attn_k_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_v_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_o_proj_weight: "f32[192, 192]", p_model_layers_0_mlp_gate_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_up_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_down_proj_weight: "f32[192, 1024]", p_model_layers_0_input_layernorm_weight: "f32[192]", p_model_layers_0_post_attention_layernorm_weight: "f32[192]", p_model_norm_weight: "f32[192]", p_lm_head_weight: "f32[32000, 192]", b_model_rotary_emb_inv_freq: "f32[48]", input_ids: "i64[s0, s1]", attention_mask: "i64[s0, s1 + s7]", position_ids: "i64[s0, s1]", past_key_values_key_cache_0: "f32[s0, 1, s7, 96]", past_key_values_value_cache_0: "f32[s0, 1, s7, 96]"):
#
sym_size_int_21: "Sym(s0)" = torch.ops.aten.sym_size.int(input_ids, 0)
sym_size_int_22: "Sym(s1)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_23: "Sym(s7)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:190 in forward, code: return F.embedding(
embedding: "f32[s0, s1, 192]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:561 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s1 + s7)" = sym_size_int_23 + sym_size_int_22
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:560 in forward, code: cache_position = torch.arange(
arange: "i64[s1]" = torch.ops.aten.arange.start(sym_size_int_23, add, device = device(type='cpu'), pin_memory = False); sym_size_int_23 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:567 in forward, code: causal_mask = self._update_causal_mask(
full: "f32[s1, s1 + s7]" = torch.ops.aten.full.default([sym_size_int_22, add], -3.4028234663852886e+38, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
triu: "f32[s1, s1 + s7]" = torch.ops.aten.triu.default(full, 1); full = None
arange_1: "i64[s1 + s7]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s1, 1]" = torch.ops.aten.reshape.default(arange, [-1, 1]); arange = None
gt: "b8[s1, s1 + s7]" = torch.ops.aten.gt.Tensor(arange_1, reshape); arange_1 = reshape = None
mul_: "f32[s1, s1 + s7]" = torch.ops.aten.mul_.Tensor(triu, gt); triu = gt = None
unsqueeze: "f32[1, s1, s1 + s7]" = torch.ops.aten.unsqueeze.default(mul_, 0); mul_ = None
unsqueeze_1: "f32[1, 1, s1, s1 + s7]" = torch.ops.aten.unsqueeze.default(unsqueeze, 1); unsqueeze = None
slice_1: "f32[1, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(unsqueeze_1, 2, 0, 9223372036854775807); unsqueeze_1 = None
slice_2: "f32[1, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_1, 3, 0, 9223372036854775807); slice_1 = None
expand: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.expand.default(slice_2, [sym_size_int_21, 1, -1, -1]); slice_2 = None
clone: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.clone.default(expand); expand = None
slice_3: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_4: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_3, 1, 0, 9223372036854775807); slice_3 = None
slice_5: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_4, 2, 0, 9223372036854775807); slice_4 = None
slice_6: "i64[s0, s1 + s7]" = torch.ops.aten.slice.Tensor(attention_mask, 0, 0, 9223372036854775807); attention_mask = None
unsqueeze_2: "i64[s0, 1, s1 + s7]" = torch.ops.aten.unsqueeze.default(slice_6, 1); slice_6 = None
unsqueeze_3: "i64[s0, 1, 1, s1 + s7]" = torch.ops.aten.unsqueeze.default(unsqueeze_2, 2); unsqueeze_2 = None
slice_7: "i64[s0, 1, 1, s1 + s7]" = torch.ops.aten.slice.Tensor(unsqueeze_3, 3, 0, 9223372036854775807); unsqueeze_3 = None
to: "i64[s0, 1, 1, s1 + s7]" = torch.ops.aten.to.dtype_layout(slice_7, dtype = torch.int64, layout = torch.strided, device = device(type='cpu')); slice_7 = None
add_2: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.add.Tensor(slice_5, to); slice_5 = to = None
eq_9: "b8[s0, 1, s1, s1 + s7]" = torch.ops.aten.eq.Scalar(add_2, 0); add_2 = None
slice_8: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_9: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_8, 1, 0, 9223372036854775807); slice_8 = None
slice_10: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_9, 2, 0, 9223372036854775807); slice_9 = None
masked_fill: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.masked_fill.Scalar(slice_10, eq_9, -3.4028234663852886e+38); slice_10 = eq_9 = None
slice_11: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_12: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_11, 1, 0, 9223372036854775807); slice_11 = None
slice_13: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_12, 2, 0, 9223372036854775807); slice_12 = None
copy_: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.copy_.default(slice_13, masked_fill); slice_13 = masked_fill = copy_ = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_set_grad_enabled = torch.ops.higher_order.wrap_with_set_grad_enabled(False, submod_3, b_model_rotary_emb_inv_freq, sym_size_int_21, position_ids); submod_3 = b_model_rotary_emb_inv_freq = position_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:148 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_6: "f32[s0, s1, 96]" = wrap_with_set_grad_enabled[0]
to_7: "f32[s0, s1, 96]" = wrap_with_set_grad_enabled[1]; wrap_with_set_grad_enabled = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:78 in forward, code: hidden_states = hidden_states.to(torch.float32)
to_8: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:79 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s0, s1, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_8, 2)
mean: "f32[s0, s1, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s0, s1, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s0, s1, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(to_8, rsqrt); rsqrt = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: return self.weight * hidden_states.to(input_dtype)
to_9: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_9); p_model_layers_0_input_layernorm_weight = to_9 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s0, s1, 192]" = 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: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:277 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s0, s1, 2, 96]" = torch.ops.aten.view.default(linear, [sym_size_int_21, sym_size_int_22, -1, 96]); linear = None
transpose_1: "f32[s0, 2, s1, 96]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s0, s1, 96]" = 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: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:278 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s0, s1, 1, 96]" = torch.ops.aten.view.default(linear_1, [sym_size_int_21, sym_size_int_22, -1, 96]); linear_1 = None
transpose_2: "f32[s0, 1, s1, 96]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s0, s1, 96]" = 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: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:279 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s0, s1, 1, 96]" = torch.ops.aten.view.default(linear_2, [sym_size_int_21, sym_size_int_22, -1, 96]); linear_2 = None
transpose_3: "f32[s0, 1, s1, 96]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:282 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_7: "f32[s0, 1, s1, 96]" = torch.ops.aten.unsqueeze.default(to_6, 1); to_6 = None
unsqueeze_8: "f32[s0, 1, s1, 96]" = torch.ops.aten.unsqueeze.default(to_7, 1); to_7 = None
mul_4: "f32[s0, 2, s1, 96]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_7)
slice_17: "f32[s0, 2, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 48)
slice_18: "f32[s0, 2, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 48, 9223372036854775807); transpose_1 = None
neg: "f32[s0, 2, s1, 48]" = torch.ops.aten.neg.default(slice_18); slice_18 = None
cat_1: "f32[s0, 2, s1, 96]" = torch.ops.aten.cat.default([neg, slice_17], -1); neg = slice_17 = None
mul_5: "f32[s0, 2, s1, 96]" = torch.ops.aten.mul.Tensor(cat_1, unsqueeze_8); cat_1 = None
add_4: "f32[s0, 2, s1, 96]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s0, 1, s1, 96]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_7); unsqueeze_7 = None
slice_19: "f32[s0, 1, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 48)
slice_20: "f32[s0, 1, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 48, 9223372036854775807); transpose_2 = None
neg_1: "f32[s0, 1, s1, 48]" = torch.ops.aten.neg.default(slice_20); slice_20 = None
cat_2: "f32[s0, 1, s1, 96]" = torch.ops.aten.cat.default([neg_1, slice_19], -1); neg_1 = slice_19 = None
mul_7: "f32[s0, 1, s1, 96]" = torch.ops.aten.mul.Tensor(cat_2, unsqueeze_8); cat_2 = unsqueeze_8 = None
add_5: "f32[s0, 1, s1, 96]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:287 in forward, code: key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_3: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.cat.default([past_key_values_key_cache_0, add_5], -2); past_key_values_key_cache_0 = add_5 = None
cat_4: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.cat.default([past_key_values_value_cache_0, transpose_3], -2); past_key_values_value_cache_0 = transpose_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:299 in forward, code: attn_output, attn_weights = attention_interface(
slice_21: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(cat_3, 0, 0, 9223372036854775807)
slice_22: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_21, 1, 0, 9223372036854775807); slice_21 = None
unsqueeze_9: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.unsqueeze.default(slice_22, 2); slice_22 = None
slice_23: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_9, 3, 0, 9223372036854775807); unsqueeze_9 = None
slice_24: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_23, 4, 0, 9223372036854775807); slice_23 = None
expand_2: "f32[s0, 1, 2, s1 + s7, 96]" = torch.ops.aten.expand.default(slice_24, [sym_size_int_21, 1, 2, add, 96]); slice_24 = None
reshape_1: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.reshape.default(expand_2, [sym_size_int_21, 2, add, 96]); expand_2 = None
slice_25: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(cat_4, 0, 0, 9223372036854775807)
slice_26: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_25, 1, 0, 9223372036854775807); slice_25 = None
unsqueeze_10: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.unsqueeze.default(slice_26, 2); slice_26 = None
slice_27: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_10, 3, 0, 9223372036854775807); unsqueeze_10 = None
slice_28: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_27, 4, 0, 9223372036854775807); slice_27 = None
expand_3: "f32[s0, 1, 2, s1 + s7, 96]" = torch.ops.aten.expand.default(slice_28, [sym_size_int_21, 1, 2, add, 96]); slice_28 = None
reshape_2: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.reshape.default(expand_3, [sym_size_int_21, 2, add, 96]); expand_3 = add = None
slice_29: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807); clone = None
slice_30: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_29, 1, 0, 9223372036854775807); slice_29 = None
slice_31: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_30, 2, 0, 9223372036854775807); slice_30 = None
contiguous: "f32[s0, 2, s1, 96]" = torch.ops.aten.contiguous.default(add_4); add_4 = None
contiguous_1: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.contiguous.default(reshape_1); reshape_1 = None
contiguous_2: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.contiguous.default(reshape_2); reshape_2 = None
scaled_dot_product_attention: "f32[s0, 2, s1, 96]" = torch.ops.aten.scaled_dot_product_attention.default(contiguous, contiguous_1, contiguous_2, slice_31, scale = 0.10206207261596575); contiguous = contiguous_1 = contiguous_2 = slice_31 = None
transpose_4: "f32[s0, s1, 2, 96]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
contiguous_3: "f32[s0, s1, 2, 96]" = torch.ops.aten.contiguous.default(transpose_4); transpose_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:310 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_3: "f32[s0, s1, 192]" = torch.ops.aten.reshape.default(contiguous_3, [sym_size_int_21, sym_size_int_22, -1]); contiguous_3 = sym_size_int_21 = sym_size_int_22 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s0, s1, 192]" = torch.ops.aten.linear.default(reshape_3, p_model_layers_0_self_attn_o_proj_weight); reshape_3 = p_model_layers_0_self_attn_o_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:354 in forward, code: hidden_states = residual + hidden_states
add_7: "f32[s0, s1, 192]" = torch.ops.aten.add.Tensor(to_8, linear_3); to_8 = linear_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:78 in forward, code: hidden_states = hidden_states.to(torch.float32)
to_10: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(add_7, torch.float32); add_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:79 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s0, s1, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_10, 2)
mean_1: "f32[s0, s1, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_8: "f32[s0, s1, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s0, s1, 1]" = torch.ops.aten.rsqrt.default(add_8); add_8 = None
mul_8: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(to_10, rsqrt_1); rsqrt_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: return self.weight * hidden_states.to(input_dtype)
to_11: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(mul_8, torch.float32); mul_8 = None
mul_9: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_11); p_model_layers_0_post_attention_layernorm_weight = to_11 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s0, s1, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:432 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s0, s1, 1024]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s0, s1, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_up_proj_weight); mul_9 = p_model_layers_0_mlp_up_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:197 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_10: "f32[s0, s1, 1024]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s0, s1, 192]" = torch.ops.aten.linear.default(mul_10, p_model_layers_0_mlp_down_proj_weight); mul_10 = p_model_layers_0_mlp_down_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:360 in forward, code: hidden_states = residual + hidden_states
add_9: "f32[s0, s1, 192]" = torch.ops.aten.add.Tensor(to_10, linear_6); to_10 = linear_6 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:78 in forward, code: hidden_states = hidden_states.to(torch.float32)
to_12: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(add_9, torch.float32); add_9 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:79 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s0, s1, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_12, 2)
mean_2: "f32[s0, s1, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_10: "f32[s0, s1, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s0, s1, 1]" = torch.ops.aten.rsqrt.default(add_10); add_10 = None
mul_11: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(to_12, rsqrt_2); to_12 = rsqrt_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: return self.weight * hidden_states.to(input_dtype)
to_13: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(mul_11, torch.float32); mul_11 = None
mul_12: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_13); p_model_norm_weight = to_13 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:866 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_32: "f32[s0, s1, 192]" = torch.ops.aten.slice.Tensor(mul_12, 0, 0, 9223372036854775807); mul_12 = None
slice_33: "f32[s0, s1, 192]" = torch.ops.aten.slice.Tensor(slice_32, 1, 0, 9223372036854775807); slice_32 = None
slice_34: "f32[s0, s1, 192]" = torch.ops.aten.slice.Tensor(slice_33, 2, 0, 9223372036854775807); slice_33 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s0, s1, 32000]" = torch.ops.aten.linear.default(slice_34, p_lm_head_weight); slice_34 = p_lm_head_weight = None
return (linear_7, cat_3, cat_4)
class submod_1(torch.nn.Module):
def forward(self, b_model_rotary_emb_inv_freq: "f32[48]", sym_size_int_21: "Sym(s0)", position_ids: "i64[s0, s1]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:133 in forward, code: inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
unsqueeze_4: "f32[1, 48]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
slice_14: "f32[1, 48]" = torch.ops.aten.slice.Tensor(unsqueeze_4, 1, 0, 9223372036854775807); unsqueeze_4 = None
unsqueeze_5: "f32[1, 48, 1]" = torch.ops.aten.unsqueeze.default(slice_14, 2); slice_14 = None
to_1: "f32[1, 48, 1]" = torch.ops.aten.to.dtype(unsqueeze_5, torch.float32); unsqueeze_5 = None
expand_1: "f32[s0, 48, 1]" = torch.ops.aten.expand.default(to_1, [sym_size_int_21, -1, 1]); to_1 = sym_size_int_21 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:134 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
slice_15: "i64[s0, s1]" = torch.ops.aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807); position_ids = None
unsqueeze_6: "i64[s0, 1, s1]" = torch.ops.aten.unsqueeze.default(slice_15, 1); slice_15 = None
slice_16: "i64[s0, 1, s1]" = torch.ops.aten.slice.Tensor(unsqueeze_6, 2, 0, 9223372036854775807); unsqueeze_6 = None
to_2: "f32[s0, 1, s1]" = torch.ops.aten.to.dtype(slice_16, torch.float32); slice_16 = 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, expand_1, to_2); submod_3 = expand_1 = to_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:141 in forward, code: cos = emb.cos()
cos: "f32[s0, s1, 96]" = wrap_with_autocast[0]
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:142 in forward, code: sin = emb.sin()
sin: "f32[s0, s1, 96]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:145 in forward, code: cos = cos * self.attention_scaling
mul: "f32[s0, s1, 96]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:146 in forward, code: sin = sin * self.attention_scaling
mul_1: "f32[s0, s1, 96]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:148 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_6: "f32[s0, s1, 96]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
to_7: "f32[s0, s1, 96]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
return (to_6, to_7)
class submod_1(torch.nn.Module):
def forward(self, expand_1: "f32[s0, 48, 1]", to_2: "f32[s0, 1, s1]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:139 in forward, code: freqs = (inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()).transpose(1, 2)
to_3: "f32[s0, 48, 1]" = torch.ops.aten.to.dtype(expand_1, torch.float32); expand_1 = None
to_4: "f32[s0, 48, 1]" = torch.ops.aten.to.dtype_layout(to_3, dtype = torch.float32, layout = torch.strided, device = device(type='cpu')); to_3 = None
to_5: "f32[s0, 1, s1]" = torch.ops.aten.to.dtype(to_2, torch.float32); to_2 = None
matmul: "f32[s0, 48, s1]" = torch.ops.aten.matmul.default(to_4, to_5); to_4 = to_5 = None
transpose: "f32[s0, s1, 48]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:140 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat: "f32[s0, s1, 96]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:141 in forward, code: cos = emb.cos()
cos: "f32[s0, s1, 96]" = torch.ops.aten.cos.default(cat)
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:142 in forward, code: sin = emb.sin()
sin: "f32[s0, s1, 96]" = torch.ops.aten.sin.default(cat); cat = None
return (cos, sin)
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_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_value_cache_0: USER_INPUT
# outputs
linear_7: USER_OUTPUT
cat_3: USER_OUTPUT
cat_4: USER_OUTPUT
Range constraints: {s0: VR[1, 1024], s1: VR[2, 4096], s1 + s7: VR[4, 8192], s7: VR[1, 4096]}
[bypass_export_some_errors] remove patches
[bypass_export_some_errors] restored sympy functions
[bypass_export_some_errors] restored pytorch functions
[bypass_export_some_errors] restored shape constraints
[unpatch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_DynamicCache: reorder_cache, update, crop, from_batch_splits, get_seq_length
[unpatch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_GenerationMixin: _cache_dependant_input_preparation, _cache_dependant_input_preparation_exporting, prepare_inputs_for_generation
[_unregister_cache_serialization] unregistered MambaCache
[_unregister_cache_serialization] unregistered DynamicCache
With the original model¶
MODEL_NAME = "arnir0/Tiny-LLM"
tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)
model = transformers.AutoModelForCausalLM.from_pretrained(MODEL_NAME)
cloned_inputs = copy.deepcopy(inputs)
with bypass_export_some_errors(patch_transformers=True, verbose=10) as modificator:
ep = torch.export.export(
model,
(),
kwargs=modificator(cloned_inputs),
dynamic_shapes=dynamic_shapes,
strict=False, # mandatory for torch==2.6
)
print("It worked:")
print(ep)
[bypass_export_some_errors] replace torch.jit.isinstance, torch._dynamo.mark_static_address
[_register_cache_serialization] register MambaCache
[_register_cache_serialization] register DynamicCache
[bypass_export_some_errors] patch sympy
[bypass_export_some_errors] patch pytorch
[bypass_export_some_errors] modifies shape constraints
[patch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_DynamicCache: reorder_cache, update, crop, from_batch_splits, get_seq_length
[patch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_GenerationMixin: _cache_dependant_input_preparation, _cache_dependant_input_preparation_exporting, prepare_inputs_for_generation
[bypass_export_some_errors] done patching
/home/xadupre/vv/this312/lib/python3.12/site-packages/torch/backends/mkldnn/__init__.py:78: UserWarning: TF32 acceleration on top of oneDNN is available for Intel GPUs. The current Torch version does not have Intel GPU Support. (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:148.)
torch._C._set_onednn_allow_tf32(_allow_tf32)
It worked:
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 192]", p_model_layers_0_self_attn_q_proj_weight: "f32[192, 192]", p_model_layers_0_self_attn_k_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_v_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_o_proj_weight: "f32[192, 192]", p_model_layers_0_mlp_gate_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_up_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_down_proj_weight: "f32[192, 1024]", p_model_layers_0_input_layernorm_weight: "f32[192]", p_model_layers_0_post_attention_layernorm_weight: "f32[192]", p_model_norm_weight: "f32[192]", p_lm_head_weight: "f32[32000, 192]", b_model_rotary_emb_inv_freq: "f32[48]", input_ids: "i64[s0, s1]", attention_mask: "i64[s0, s1 + s7]", position_ids: "i64[s0, s1]", past_key_values_key_cache_0: "f32[s0, 1, s7, 96]", past_key_values_value_cache_0: "f32[s0, 1, s7, 96]"):
#
sym_size_int_21: "Sym(s0)" = torch.ops.aten.sym_size.int(input_ids, 0)
sym_size_int_22: "Sym(s1)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_23: "Sym(s7)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:190 in forward, code: return F.embedding(
embedding: "f32[s0, s1, 192]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:561 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s1 + s7)" = sym_size_int_23 + sym_size_int_22
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:560 in forward, code: cache_position = torch.arange(
arange: "i64[s1]" = torch.ops.aten.arange.start(sym_size_int_23, add, device = device(type='cpu'), pin_memory = False); sym_size_int_23 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:567 in forward, code: causal_mask = self._update_causal_mask(
full: "f32[s1, s1 + s7]" = torch.ops.aten.full.default([sym_size_int_22, add], -3.4028234663852886e+38, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
triu: "f32[s1, s1 + s7]" = torch.ops.aten.triu.default(full, 1); full = None
arange_1: "i64[s1 + s7]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s1, 1]" = torch.ops.aten.reshape.default(arange, [-1, 1]); arange = None
gt: "b8[s1, s1 + s7]" = torch.ops.aten.gt.Tensor(arange_1, reshape); arange_1 = reshape = None
mul_: "f32[s1, s1 + s7]" = torch.ops.aten.mul_.Tensor(triu, gt); triu = gt = None
unsqueeze: "f32[1, s1, s1 + s7]" = torch.ops.aten.unsqueeze.default(mul_, 0); mul_ = None
unsqueeze_1: "f32[1, 1, s1, s1 + s7]" = torch.ops.aten.unsqueeze.default(unsqueeze, 1); unsqueeze = None
slice_1: "f32[1, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(unsqueeze_1, 2, 0, 9223372036854775807); unsqueeze_1 = None
slice_2: "f32[1, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_1, 3, 0, 9223372036854775807); slice_1 = None
expand: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.expand.default(slice_2, [sym_size_int_21, 1, -1, -1]); slice_2 = None
clone: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.clone.default(expand); expand = None
slice_3: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_4: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_3, 1, 0, 9223372036854775807); slice_3 = None
slice_5: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_4, 2, 0, 9223372036854775807); slice_4 = None
slice_6: "i64[s0, s1 + s7]" = torch.ops.aten.slice.Tensor(attention_mask, 0, 0, 9223372036854775807); attention_mask = None
unsqueeze_2: "i64[s0, 1, s1 + s7]" = torch.ops.aten.unsqueeze.default(slice_6, 1); slice_6 = None
unsqueeze_3: "i64[s0, 1, 1, s1 + s7]" = torch.ops.aten.unsqueeze.default(unsqueeze_2, 2); unsqueeze_2 = None
slice_7: "i64[s0, 1, 1, s1 + s7]" = torch.ops.aten.slice.Tensor(unsqueeze_3, 3, 0, 9223372036854775807); unsqueeze_3 = None
to: "i64[s0, 1, 1, s1 + s7]" = torch.ops.aten.to.dtype_layout(slice_7, dtype = torch.int64, layout = torch.strided, device = device(type='cpu')); slice_7 = None
add_2: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.add.Tensor(slice_5, to); slice_5 = to = None
eq_9: "b8[s0, 1, s1, s1 + s7]" = torch.ops.aten.eq.Scalar(add_2, 0); add_2 = None
slice_8: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_9: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_8, 1, 0, 9223372036854775807); slice_8 = None
slice_10: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_9, 2, 0, 9223372036854775807); slice_9 = None
masked_fill: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.masked_fill.Scalar(slice_10, eq_9, -3.4028234663852886e+38); slice_10 = eq_9 = None
slice_11: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_12: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_11, 1, 0, 9223372036854775807); slice_11 = None
slice_13: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_12, 2, 0, 9223372036854775807); slice_12 = None
copy_: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.copy_.default(slice_13, masked_fill); slice_13 = masked_fill = copy_ = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_set_grad_enabled = torch.ops.higher_order.wrap_with_set_grad_enabled(False, submod_3, b_model_rotary_emb_inv_freq, sym_size_int_21, position_ids); submod_3 = b_model_rotary_emb_inv_freq = position_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:148 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_6: "f32[s0, s1, 96]" = wrap_with_set_grad_enabled[0]
to_7: "f32[s0, s1, 96]" = wrap_with_set_grad_enabled[1]; wrap_with_set_grad_enabled = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:78 in forward, code: hidden_states = hidden_states.to(torch.float32)
to_8: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:79 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s0, s1, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_8, 2)
mean: "f32[s0, s1, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s0, s1, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s0, s1, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(to_8, rsqrt); rsqrt = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: return self.weight * hidden_states.to(input_dtype)
to_9: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_9); p_model_layers_0_input_layernorm_weight = to_9 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s0, s1, 192]" = 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: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:277 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s0, s1, 2, 96]" = torch.ops.aten.view.default(linear, [sym_size_int_21, sym_size_int_22, -1, 96]); linear = None
transpose_1: "f32[s0, 2, s1, 96]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s0, s1, 96]" = 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: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:278 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s0, s1, 1, 96]" = torch.ops.aten.view.default(linear_1, [sym_size_int_21, sym_size_int_22, -1, 96]); linear_1 = None
transpose_2: "f32[s0, 1, s1, 96]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s0, s1, 96]" = 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: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:279 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s0, s1, 1, 96]" = torch.ops.aten.view.default(linear_2, [sym_size_int_21, sym_size_int_22, -1, 96]); linear_2 = None
transpose_3: "f32[s0, 1, s1, 96]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:282 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_7: "f32[s0, 1, s1, 96]" = torch.ops.aten.unsqueeze.default(to_6, 1); to_6 = None
unsqueeze_8: "f32[s0, 1, s1, 96]" = torch.ops.aten.unsqueeze.default(to_7, 1); to_7 = None
mul_4: "f32[s0, 2, s1, 96]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_7)
slice_17: "f32[s0, 2, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 48)
slice_18: "f32[s0, 2, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 48, 9223372036854775807); transpose_1 = None
neg: "f32[s0, 2, s1, 48]" = torch.ops.aten.neg.default(slice_18); slice_18 = None
cat_1: "f32[s0, 2, s1, 96]" = torch.ops.aten.cat.default([neg, slice_17], -1); neg = slice_17 = None
mul_5: "f32[s0, 2, s1, 96]" = torch.ops.aten.mul.Tensor(cat_1, unsqueeze_8); cat_1 = None
add_4: "f32[s0, 2, s1, 96]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s0, 1, s1, 96]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_7); unsqueeze_7 = None
slice_19: "f32[s0, 1, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 48)
slice_20: "f32[s0, 1, s1, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 48, 9223372036854775807); transpose_2 = None
neg_1: "f32[s0, 1, s1, 48]" = torch.ops.aten.neg.default(slice_20); slice_20 = None
cat_2: "f32[s0, 1, s1, 96]" = torch.ops.aten.cat.default([neg_1, slice_19], -1); neg_1 = slice_19 = None
mul_7: "f32[s0, 1, s1, 96]" = torch.ops.aten.mul.Tensor(cat_2, unsqueeze_8); cat_2 = unsqueeze_8 = None
add_5: "f32[s0, 1, s1, 96]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:287 in forward, code: key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_3: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.cat.default([past_key_values_key_cache_0, add_5], -2); past_key_values_key_cache_0 = add_5 = None
cat_4: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.cat.default([past_key_values_value_cache_0, transpose_3], -2); past_key_values_value_cache_0 = transpose_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:299 in forward, code: attn_output, attn_weights = attention_interface(
slice_21: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(cat_3, 0, 0, 9223372036854775807)
slice_22: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_21, 1, 0, 9223372036854775807); slice_21 = None
unsqueeze_9: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.unsqueeze.default(slice_22, 2); slice_22 = None
slice_23: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_9, 3, 0, 9223372036854775807); unsqueeze_9 = None
slice_24: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_23, 4, 0, 9223372036854775807); slice_23 = None
expand_2: "f32[s0, 1, 2, s1 + s7, 96]" = torch.ops.aten.expand.default(slice_24, [sym_size_int_21, 1, 2, add, 96]); slice_24 = None
reshape_1: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.reshape.default(expand_2, [sym_size_int_21, 2, add, 96]); expand_2 = None
slice_25: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(cat_4, 0, 0, 9223372036854775807)
slice_26: "f32[s0, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_25, 1, 0, 9223372036854775807); slice_25 = None
unsqueeze_10: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.unsqueeze.default(slice_26, 2); slice_26 = None
slice_27: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_10, 3, 0, 9223372036854775807); unsqueeze_10 = None
slice_28: "f32[s0, 1, 1, s1 + s7, 96]" = torch.ops.aten.slice.Tensor(slice_27, 4, 0, 9223372036854775807); slice_27 = None
expand_3: "f32[s0, 1, 2, s1 + s7, 96]" = torch.ops.aten.expand.default(slice_28, [sym_size_int_21, 1, 2, add, 96]); slice_28 = None
reshape_2: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.reshape.default(expand_3, [sym_size_int_21, 2, add, 96]); expand_3 = add = None
slice_29: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807); clone = None
slice_30: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_29, 1, 0, 9223372036854775807); slice_29 = None
slice_31: "f32[s0, 1, s1, s1 + s7]" = torch.ops.aten.slice.Tensor(slice_30, 2, 0, 9223372036854775807); slice_30 = None
contiguous: "f32[s0, 2, s1, 96]" = torch.ops.aten.contiguous.default(add_4); add_4 = None
contiguous_1: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.contiguous.default(reshape_1); reshape_1 = None
contiguous_2: "f32[s0, 2, s1 + s7, 96]" = torch.ops.aten.contiguous.default(reshape_2); reshape_2 = None
scaled_dot_product_attention: "f32[s0, 2, s1, 96]" = torch.ops.aten.scaled_dot_product_attention.default(contiguous, contiguous_1, contiguous_2, slice_31, scale = 0.10206207261596575); contiguous = contiguous_1 = contiguous_2 = slice_31 = None
transpose_4: "f32[s0, s1, 2, 96]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
contiguous_3: "f32[s0, s1, 2, 96]" = torch.ops.aten.contiguous.default(transpose_4); transpose_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:310 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_3: "f32[s0, s1, 192]" = torch.ops.aten.reshape.default(contiguous_3, [sym_size_int_21, sym_size_int_22, -1]); contiguous_3 = sym_size_int_21 = sym_size_int_22 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s0, s1, 192]" = torch.ops.aten.linear.default(reshape_3, p_model_layers_0_self_attn_o_proj_weight); reshape_3 = p_model_layers_0_self_attn_o_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:354 in forward, code: hidden_states = residual + hidden_states
add_7: "f32[s0, s1, 192]" = torch.ops.aten.add.Tensor(to_8, linear_3); to_8 = linear_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:78 in forward, code: hidden_states = hidden_states.to(torch.float32)
to_10: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(add_7, torch.float32); add_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:79 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s0, s1, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_10, 2)
mean_1: "f32[s0, s1, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_8: "f32[s0, s1, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s0, s1, 1]" = torch.ops.aten.rsqrt.default(add_8); add_8 = None
mul_8: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(to_10, rsqrt_1); rsqrt_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: return self.weight * hidden_states.to(input_dtype)
to_11: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(mul_8, torch.float32); mul_8 = None
mul_9: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_11); p_model_layers_0_post_attention_layernorm_weight = to_11 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s0, s1, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:432 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s0, s1, 1024]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s0, s1, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_up_proj_weight); mul_9 = p_model_layers_0_mlp_up_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:197 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_10: "f32[s0, s1, 1024]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s0, s1, 192]" = torch.ops.aten.linear.default(mul_10, p_model_layers_0_mlp_down_proj_weight); mul_10 = p_model_layers_0_mlp_down_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:360 in forward, code: hidden_states = residual + hidden_states
add_9: "f32[s0, s1, 192]" = torch.ops.aten.add.Tensor(to_10, linear_6); to_10 = linear_6 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:78 in forward, code: hidden_states = hidden_states.to(torch.float32)
to_12: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(add_9, torch.float32); add_9 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:79 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s0, s1, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_12, 2)
mean_2: "f32[s0, s1, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:80 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_10: "f32[s0, s1, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s0, s1, 1]" = torch.ops.aten.rsqrt.default(add_10); add_10 = None
mul_11: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(to_12, rsqrt_2); to_12 = rsqrt_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:81 in forward, code: return self.weight * hidden_states.to(input_dtype)
to_13: "f32[s0, s1, 192]" = torch.ops.aten.to.dtype(mul_11, torch.float32); mul_11 = None
mul_12: "f32[s0, s1, 192]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_13); p_model_norm_weight = to_13 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:866 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_32: "f32[s0, s1, 192]" = torch.ops.aten.slice.Tensor(mul_12, 0, 0, 9223372036854775807); mul_12 = None
slice_33: "f32[s0, s1, 192]" = torch.ops.aten.slice.Tensor(slice_32, 1, 0, 9223372036854775807); slice_32 = None
slice_34: "f32[s0, s1, 192]" = torch.ops.aten.slice.Tensor(slice_33, 2, 0, 9223372036854775807); slice_33 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s0, s1, 32000]" = torch.ops.aten.linear.default(slice_34, p_lm_head_weight); slice_34 = p_lm_head_weight = None
return (linear_7, cat_3, cat_4)
class submod_1(torch.nn.Module):
def forward(self, b_model_rotary_emb_inv_freq: "f32[48]", sym_size_int_21: "Sym(s0)", position_ids: "i64[s0, s1]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:133 in forward, code: inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
unsqueeze_4: "f32[1, 48]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
slice_14: "f32[1, 48]" = torch.ops.aten.slice.Tensor(unsqueeze_4, 1, 0, 9223372036854775807); unsqueeze_4 = None
unsqueeze_5: "f32[1, 48, 1]" = torch.ops.aten.unsqueeze.default(slice_14, 2); slice_14 = None
to_1: "f32[1, 48, 1]" = torch.ops.aten.to.dtype(unsqueeze_5, torch.float32); unsqueeze_5 = None
expand_1: "f32[s0, 48, 1]" = torch.ops.aten.expand.default(to_1, [sym_size_int_21, -1, 1]); to_1 = sym_size_int_21 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:134 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
slice_15: "i64[s0, s1]" = torch.ops.aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807); position_ids = None
unsqueeze_6: "i64[s0, 1, s1]" = torch.ops.aten.unsqueeze.default(slice_15, 1); slice_15 = None
slice_16: "i64[s0, 1, s1]" = torch.ops.aten.slice.Tensor(unsqueeze_6, 2, 0, 9223372036854775807); unsqueeze_6 = None
to_2: "f32[s0, 1, s1]" = torch.ops.aten.to.dtype(slice_16, torch.float32); slice_16 = 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, expand_1, to_2); submod_3 = expand_1 = to_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:141 in forward, code: cos = emb.cos()
cos: "f32[s0, s1, 96]" = wrap_with_autocast[0]
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:142 in forward, code: sin = emb.sin()
sin: "f32[s0, s1, 96]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:145 in forward, code: cos = cos * self.attention_scaling
mul: "f32[s0, s1, 96]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:146 in forward, code: sin = sin * self.attention_scaling
mul_1: "f32[s0, s1, 96]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:148 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_6: "f32[s0, s1, 96]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
to_7: "f32[s0, s1, 96]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
return (to_6, to_7)
class submod_1(torch.nn.Module):
def forward(self, expand_1: "f32[s0, 48, 1]", to_2: "f32[s0, 1, s1]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:139 in forward, code: freqs = (inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()).transpose(1, 2)
to_3: "f32[s0, 48, 1]" = torch.ops.aten.to.dtype(expand_1, torch.float32); expand_1 = None
to_4: "f32[s0, 48, 1]" = torch.ops.aten.to.dtype_layout(to_3, dtype = torch.float32, layout = torch.strided, device = device(type='cpu')); to_3 = None
to_5: "f32[s0, 1, s1]" = torch.ops.aten.to.dtype(to_2, torch.float32); to_2 = None
matmul: "f32[s0, 48, s1]" = torch.ops.aten.matmul.default(to_4, to_5); to_4 = to_5 = None
transpose: "f32[s0, s1, 48]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:140 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat: "f32[s0, s1, 96]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:141 in forward, code: cos = emb.cos()
cos: "f32[s0, s1, 96]" = torch.ops.aten.cos.default(cat)
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:142 in forward, code: sin = emb.sin()
sin: "f32[s0, s1, 96]" = torch.ops.aten.sin.default(cat); cat = None
return (cos, sin)
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_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_value_cache_0: USER_INPUT
# outputs
linear_7: USER_OUTPUT
cat_3: USER_OUTPUT
cat_4: USER_OUTPUT
Range constraints: {s0: VR[1, 1024], s1: VR[2, 4096], s1 + s7: VR[4, 8192], s7: VR[1, 4096]}
[bypass_export_some_errors] remove patches
[bypass_export_some_errors] restored sympy functions
[bypass_export_some_errors] restored pytorch functions
[bypass_export_some_errors] restored shape constraints
[unpatch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_DynamicCache: reorder_cache, update, crop, from_batch_splits, get_seq_length
[unpatch_module] onnx_diagnostic.torch_export_patches.patches.patch_transformers - patched_GenerationMixin: _cache_dependant_input_preparation, _cache_dependant_input_preparation_exporting, prepare_inputs_for_generation
[_unregister_cache_serialization] unregistered MambaCache
[_unregister_cache_serialization] unregistered DynamicCache
doc.plot_legend("Tiny-LLM patched", "torch.export.export", "green")
Total running time of the script: (0 minutes 12.619 seconds)
Related examples
Steel method forward to guess the dynamic shapes (with Tiny-LLM)
Steel method forward to guess the dynamic shapes (with Tiny-LLM)
Find and fix an export issue due to dynamic shapes
Find and fix an export issue due to dynamic shapes