Llama 3.2 vision finetuning error (Unsupported: hasattr ConstDictVariable to) #1325

adi7820 · 2024-11-22T13:38:27Z

Unsupported Traceback (most recent call last)
Cell In[10], line 22
20 from transformers import TextStreamer
21 text_streamer = TextStreamer(tokenizer, skip_prompt = True)
---> 22 _ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128,
23 use_cache = True, temperature = 1.5, min_p = 0.1)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/utils/_contextlib.py:116, in context_decorator..decorate_context(*args, **kwargs)
113 @functools.wraps(func)
114 def decorate_context(*args, **kwargs):
115 with ctx_factory():
--> 116 return func(*args, **kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/unsloth/models/vision.py:63, in _wrap_fast_inference.._fast_generate(*args, **kwargs)
61 # Autocasted
62 with torch.autocast(device_type = device_type, dtype = dtype):
---> 63 output = generate(*args, **kwargs)
64 pass
65 return output

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/peft/peft_model.py:1704, in PeftModelForCausalLM.generate(self, *args, **kwargs)
1702 with self._enable_peft_forward_hooks(*args, **kwargs):
1703 kwargs = {k: v for k, v in kwargs.items() if k not in self.special_peft_forward_args}
-> 1704 outputs = self.base_model.generate(*args, **kwargs)
1705 else:
1706 outputs = self.base_model.generate(**kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/utils/_contextlib.py:116, in context_decorator..decorate_context(*args, **kwargs)
113 @functools.wraps(func)
114 def decorate_context(*args, **kwargs):
115 with ctx_factory():
--> 116 return func(*args, **kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/transformers/generation/utils.py:2215, in GenerationMixin.generate(self, inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, assistant_model, streamer, negative_prompt_ids, negative_prompt_attention_mask, **kwargs)
2207 input_ids, model_kwargs = self._expand_inputs_for_generation(
2208 input_ids=input_ids,
2209 expand_size=generation_config.num_return_sequences,
2210 is_encoder_decoder=self.config.is_encoder_decoder,
2211 **model_kwargs,
2212 )
2214 # 12. run sample (it degenerates to greedy search when generation_config.do_sample=False)
-> 2215 result = self._sample(
2216 input_ids,
2217 logits_processor=prepared_logits_processor,
2218 stopping_criteria=prepared_stopping_criteria,
2219 generation_config=generation_config,
2220 synced_gpus=synced_gpus,
2221 streamer=streamer,
2222 **model_kwargs,
2223 )
2225 elif generation_mode in (GenerationMode.BEAM_SAMPLE, GenerationMode.BEAM_SEARCH):
2226 # 11. prepare beam search scorer
2227 beam_scorer = BeamSearchScorer(
2228 batch_size=batch_size,
2229 num_beams=generation_config.num_beams,
(...)
2234 max_length=generation_config.max_length,
2235 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/transformers/generation/utils.py:3206, in GenerationMixin._sample(self, input_ids, logits_processor, stopping_criteria, generation_config, synced_gpus, streamer, **model_kwargs)
3203 model_inputs.update({"output_hidden_states": output_hidden_states} if output_hidden_states else {})
3205 # forward pass to get next token
-> 3206 outputs = self(**model_inputs, return_dict=True)
3208 # synced_gpus: don't waste resources running the code we don't need; kwargs must be updated before skipping
3209 model_kwargs = self._update_model_kwargs_for_generation(
3210 outputs,
3211 model_kwargs,
3212 is_encoder_decoder=self.config.is_encoder_decoder,
3213 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/nn/modules/module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
1551 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1552 else:
-> 1553 return self._call_impl(*args, **kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/nn/modules/module.py:1562, in Module._call_impl(self, *args, **kwargs)
1557 # If we don't have any hooks, we want to skip the rest of the logic in
1558 # this function, and just call forward.
1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1560 or _global_backward_pre_hooks or _global_backward_hooks
1561 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1562 return forward_call(*args, **kwargs)
1564 try:
1565 result = None

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/accelerate/hooks.py:170, in add_hook_to_module..new_forward(module, *args, **kwargs)
168 output = module._old_forward(*args, **kwargs)
169 else:
--> 170 output = module._old_forward(*args, **kwargs)
171 return module._hf_hook.post_forward(module, output)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/transformers/models/mllama/modeling_mllama.py:2098, in MllamaForConditionalGeneration.forward(self, input_ids, pixel_values, aspect_ratio_mask, aspect_ratio_ids, attention_mask, cross_attention_mask, cross_attention_states, position_ids, past_key_values, inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict, cache_position, num_logits_to_keep)
2096 raise ValueError("aspect_ratio_ids must be provided if pixel_values is provided")
2097 # get vision tokens from vision model
-> 2098 vision_outputs = self.vision_model(
2099 pixel_values=pixel_values,
2100 aspect_ratio_ids=aspect_ratio_ids,
2101 aspect_ratio_mask=aspect_ratio_mask,
2102 output_hidden_states=output_hidden_states,
2103 output_attentions=output_attentions,
2104 return_dict=return_dict,
2105 )
2106 cross_attention_states = vision_outputs[0]
2107 cross_attention_states = self.multi_modal_projector(cross_attention_states).reshape(
2108 -1, cross_attention_states.shape[-2], self.hidden_size
2109 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/nn/modules/module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
1551 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1552 else:
-> 1553 return self._call_impl(*args, **kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/nn/modules/module.py:1562, in Module._call_impl(self, *args, **kwargs)
1557 # If we don't have any hooks, we want to skip the rest of the logic in
1558 # this function, and just call forward.
1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1560 or _global_backward_pre_hooks or _global_backward_hooks
1561 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1562 return forward_call(*args, **kwargs)
1564 try:
1565 result = None

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/accelerate/hooks.py:170, in add_hook_to_module..new_forward(module, *args, **kwargs)
168 output = module._old_forward(*args, **kwargs)
169 else:
--> 170 output = module._old_forward(*args, **kwargs)
171 return module._hf_hook.post_forward(module, output)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/transformers/models/mllama/modeling_mllama.py:1422, in MllamaVisionModel.forward(self, pixel_values, aspect_ratio_ids, aspect_ratio_mask, output_attentions, output_hidden_states, return_dict)
1420 _, num_patches, dim = hidden_state.shape
1421 hidden_state = hidden_state.reshape(batch_size * num_concurrent_media, num_tiles, -1, dim)
-> 1422 hidden_state = self.pre_tile_positional_embedding(hidden_state, aspect_ratio_ids)
1424 # Add cls token
1425 hidden_state = hidden_state.reshape(batch_size * num_concurrent_media * num_tiles, num_patches, dim)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/nn/modules/module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
1551 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1552 else:
-> 1553 return self._call_impl(*args, **kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/nn/modules/module.py:1562, in Module._call_impl(self, *args, **kwargs)
1557 # If we don't have any hooks, we want to skip the rest of the logic in
1558 # this function, and just call forward.
1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1560 or _global_backward_pre_hooks or _global_backward_hooks
1561 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1562 return forward_call(*args, **kwargs)
1564 try:
1565 result = None

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/accelerate/hooks.py:170, in add_hook_to_module..new_forward(module, *args, **kwargs)
168 output = module._old_forward(*args, **kwargs)
169 else:
--> 170 output = module._old_forward(*args, **kwargs)
171 return module._hf_hook.post_forward(module, output)

File ~/SageMaker/vlm-unsloth-finetune/unsloth_compiled_cache/unsloth_compiled_module_mllama.py:105, in MllamaPrecomputedAspectRatioEmbedding.forward(self, hidden_state, aspect_ratio_ids)
104 def forward(self, hidden_state: torch.Tensor, aspect_ratio_ids: torch.Tensor) -> torch.Tensor:
--> 105 return MllamaPrecomputedAspectRatioEmbedding_forward(self, hidden_state, aspect_ratio_ids)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py:433, in _TorchDynamoContext.call.._fn(*args, **kwargs)
428 saved_dynamic_layer_stack_depth = (
429 torch._C._functorch.get_dynamic_layer_stack_depth()
430 )
432 try:
--> 433 return fn(*args, **kwargs)
434 finally:
435 # Restore the dynamic layer stack depth if necessary.
436 torch._C._functorch.pop_dynamic_layer_stack_and_undo_to_depth(
437 saved_dynamic_layer_stack_depth
438 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:1116, in CatchErrorsWrapper.call(self, frame, cache_entry, frame_state)
1110 return hijacked_callback(
1111 frame, cache_entry, self.hooks, frame_state
1112 )
1114 with compile_lock, _disable_current_modes():
1115 # skip=1: skip this frame
-> 1116 return self._torchdynamo_orig_callable(
1117 frame, cache_entry, self.hooks, frame_state, skip=1
1118 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:472, in ConvertFrameAssert.call(self, frame, cache_entry, hooks, frame_state, skip)
458 compile_id = CompileId(frame_id, frame_compile_id)
460 signpost_event(
461 "dynamo",
462 "_convert_frame_assert._compile",
(...)
469 },
470 )
--> 472 return _compile(
473 frame.f_code,
474 frame.f_globals,
475 frame.f_locals,
476 frame.f_builtins,
477 self._torchdynamo_orig_callable,
478 self._one_graph,
479 self._export,
480 self._export_constraints,
481 hooks,
482 cache_entry,
483 cache_size,
484 frame,
485 frame_state=frame_state,
486 compile_id=compile_id,
487 skip=skip + 1,
488 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_utils_internal.py:84, in compile_time_strobelight_meta..compile_time_strobelight_meta_inner..wrapper_function(*args, **kwargs)
82 if "skip" in kwargs:
83 kwargs["skip"] = kwargs["skip"] + 1
---> 84 return StrobelightCompileTimeProfiler.profile_compile_time(
85 function, phase_name, *args, **kwargs
86 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_strobelight/compile_time_profiler.py:129, in StrobelightCompileTimeProfiler.profile_compile_time(cls, func, phase_name, *args, **kwargs)
124 @classmethod
125 def profile_compile_time(
126 cls, func: Any, phase_name: str, *args: Any, **kwargs: Any
127 ) -> Any:
128 if not cls.enabled:
--> 129 return func(*args, **kwargs)
131 if cls.profiler is None:
132 logger.error("profiler is not set")

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/contextlib.py:79, in ContextDecorator.call..inner(*args, **kwds)
76 @wraps(func)
77 def inner(*args, **kwds):
78 with self._recreate_cm():
---> 79 return func(*args, **kwds)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:817, in _compile(code, globals, locals, builtins, compiler_fn, one_graph, export, export_constraints, hooks, cache_entry, cache_size, frame, frame_state, compile_id, skip)
815 guarded_code = None
816 try:
--> 817 guarded_code = compile_inner(code, one_graph, hooks, transform)
818 return guarded_code
819 except (
820 Unsupported,
821 TorchRuntimeError,
(...)
828 BisectValidationException,
829 ) as e:

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/utils.py:231, in dynamo_timed..dynamo_timed_inner..time_wrapper(*args, **kwargs)
229 with torch.profiler.record_function(f"{key} (dynamo_timed)"):
230 t0 = time.time()
--> 231 r = func(*args, **kwargs)
232 time_spent = time.time() - t0
233 compilation_time_metrics[key].append(time_spent)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:636, in _compile..compile_inner(code, one_graph, hooks, transform)
634 CompileContext.get().attempt = attempt
635 try:
--> 636 out_code = transform_code_object(code, transform)
637 break
638 except exc.RestartAnalysis as e:

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/bytecode_transformation.py:1185, in transform_code_object(code, transformations, safe)
1182 instructions = cleaned_instructions(code, safe)
1183 propagate_line_nums(instructions)
-> 1185 transformations(instructions, code_options)
1186 return clean_and_assemble_instructions(instructions, keys, code_options)[1]

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:178, in preserve_global_state.._fn(*args, **kwargs)
176 cleanup = setup_compile_debug()
177 try:
--> 178 return fn(*args, **kwargs)
179 finally:
180 cleanup.close()

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:582, in _compile..transform(instructions, code_options)
580 try:
581 with tracing(tracer.output.tracing_context), tracer.set_current_tx():
--> 582 tracer.run()
583 except exc.UnspecializeRestartAnalysis:
584 speculation_log.clear()

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:2451, in InstructionTranslator.run(self)
2450 def run(self):
-> 2451 super().run()

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:893, in InstructionTranslatorBase.run(self)
891 try:
892 self.output.push_tx(self)
--> 893 while self.step():
894 pass
895 except BackendCompilerFailed:

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:805, in InstructionTranslatorBase.step(self)
802 self.update_block_stack(inst)
804 try:
--> 805 self.dispatch_table[inst.opcode](self, inst)
806 return not self.output.should_exit
807 except exc.ObservedException:

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:499, in break_graph_if_unsupported..decorator..wrapper(self, inst)
497 return handle_graph_break(self, inst, speculation.reason)
498 try:
--> 499 return inner_fn(self, inst)
500 except Unsupported as excp:
501 if self.generic_context_manager_depth > 0:
502 # We don't support graph break under GenericContextWrappingVariable,
503 # If there is, we roll back to the checkpoint and fall back.

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1459, in InstructionTranslatorBase.CALL_FUNCTION(self, inst)
1457 args = self.popn(inst.argval)
1458 fn = self.pop()
-> 1459 self.call_function(fn, args, {})

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:743, in InstructionTranslatorBase.call_function(self, fn, args, kwargs)
741 if inner_fn and callable(inner_fn) and is_forbidden(inner_fn):
742 raise AssertionError(f"Attempt to trace forbidden callable {inner_fn}")
--> 743 self.push(fn.call_function(self, args, kwargs))

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/variables/nn_module.py:838, in UnspecializedNNModuleVariable.call_function(self, tx, args, kwargs)
832 ctx = (
833 record_nn_module_stack(str(id(mod)), self.source, tx, mod)
834 if self.source
835 else nullcontext()
836 )
837 with ctx:
--> 838 return variables.UserFunctionVariable(fn, source=source).call_function(
839 tx, [self] + list(args), kwargs
840 )

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py:293, in UserFunctionVariable.call_function(self, tx, args, kwargs)
288 if self.is_constant:
289 return invoke_and_store_as_constant(
290 tx, self.fn, self.get_name(), args, kwargs
291 )
--> 293 return super().call_function(tx, args, kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py:90, in BaseUserFunctionVariable.call_function(self, tx, args, kwargs)
87 def call_function(
88 self, tx, args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]"
89 ) -> "VariableTracker":
---> 90 return tx.inline_user_function_return(self, [*self.self_args(), *args], kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:749, in InstructionTranslatorBase.inline_user_function_return(self, fn, args, kwargs)
745 def inline_user_function_return(self, fn, args, kwargs):
746 """
747 A call to some user defined function by inlining it.
748 """
--> 749 return InliningInstructionTranslator.inline_call(self, fn, args, kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/dynamo/symbolic_convert.py:2666, in InliningInstructionTranslator.inline_call(cls, parent, func, args, kwargs)
2663 @classmethod
2664 def inline_call(cls, parent, func, args, kwargs):
2665 with patch.dict(counters, {"unimplemented": counters["inline_call"]}):
-> 2666 return cls.inline_call(parent, func, args, kwargs)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/dynamo/symbolic_convert.py:2782, in InliningInstructionTranslator.inline_call(parent, func, args, kwargs)
2780 try:
2781 with strict_ctx:
-> 2782 tracer.run()
2783 except exc.ObservedException as e:
2784 msg = f"Observed exception DURING INLING {code} : {e}"

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:893, in InstructionTranslatorBase.run(self)
891 try:
892 self.output.push_tx(self)
--> 893 while self.step():
894 pass
895 except BackendCompilerFailed:

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:805, in InstructionTranslatorBase.step(self)
802 self.update_block_stack(inst)
804 try:
--> 805 self.dispatch_table[inst.opcode](self, inst)
806 return not self.output.should_exit
807 except exc.ObservedException:

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:499, in break_graph_if_unsupported..decorator..wrapper(self, inst)
497 return handle_graph_break(self, inst, speculation.reason)
498 try:
--> 499 return inner_fn(self, inst)
500 except Unsupported as excp:
501 if self.generic_context_manager_depth > 0:
502 # We don't support graph break under GenericContextWrappingVariable,
503 # If there is, we roll back to the checkpoint and fall back.

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1500, in InstructionTranslatorBase.CALL_FUNCTION_EX(self, inst)
1498 # Map to a dictionary of str -> VariableTracker
1499 kwargsvars = kwargsvars.keys_as_python_constant()
-> 1500 self.call_function(fn, argsvars.items, kwargsvars)

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:743, in InstructionTranslatorBase.call_function(self, fn, args, kwargs)
741 if inner_fn and callable(inner_fn) and is_forbidden(inner_fn):
742 raise AssertionError(f"Attempt to trace forbidden callable {inner_fn}")
--> 743 self.push(fn.call_function(self, args, kwargs))

File ~/anaconda3/envs/tensorflow2_p310/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py:796, in FunctoolsPartialVariable.call_function(self, tx, args, kwargs)
794 merged_args = self.args + args
795 merged_kwargs = {**self.keywords, **kwargs}
--> 796 return self.func.call_function(tx, merged_args, merged_kwargs)