Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[HPU] rlhf #303

Open
Delaunay opened this issue Oct 8, 2024 · 4 comments
Open

[HPU] rlhf #303

Delaunay opened this issue Oct 8, 2024 · 4 comments
Labels
HPU

Comments

@Delaunay
Copy link
Collaborator

Delaunay commented Oct 8, 2024 •
edited
Loading

Lazy Mode

habana_frameworks/torch/gpu_migration/torch/cuda/random.py

Changed

          	pyhlml.hlmlInit()
                count = pyhlml.hlmlDeviceGetCount()
          	pyhlml.hlmlShutdown()

to

        	try:
            	   pyhlml.hlmlInit()
            	   count = pyhlml.hlmlDeviceGetCount()
            	   pyhlml.hlmlShutdown()
        	except Exception:
            	  count = pyhlml.hlmlDeviceGetCount()

Latest error

rlhf-single.D5
==============
  * no training rate retrieved
  * Error codes = 1
  * 1 exceptions found
  0%|          | 0/254 [00:02<?, ?it/s]
        | Traceback (most recent call last):
        |   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 137, in <module>
        |     main()
        |   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 128, in main
        |     trainer.train()
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/ppov2_trainer.py", line 309, in train
        |     query_responses, logitss = batch_generation(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        |     return func(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1176, in batch_generation
        |     query_response, logits = generate(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1151, in generate
        |     output = lm_backbone.generate(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        |     return func(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 2024, in generate
        |     result = self._sample(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 3038, in _sample
        |     unfinished_sequences = unfinished_sequences & ~stopping_criteria(input_ids, scores)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/stopping_criteria.py", line 509, in __call__
        |     is_done = torch.full((input_ids.shape[0],), False, device=input_ids.device, dtype=torch.bool)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/habana_frameworks/torch/gpu_migration/torch/_C/_VariableFunctions.py", line 244, in full
        |     return TorchModuleRegister.full(*args, **kwargs)
        | RuntimeError: [Rank:0] FATAL ERROR :: MODULE:PT_BRIDGE Exception in acc thread pool task has been thrown: Got a non-HPU tensor, expecting an HPU tensor
        | [Rank:0] Habana exception raised from GetHbLazyTensor at aten_lazy_bridge.cpp:290
@Delaunay
Copy link
Collaborator Author

Delaunay commented Oct 9, 2024 •
edited
Loading

Changed voir so

@contextmanager
def hlmlinit():
    try:
        pyhlml.hlmlInit()
        yield
        pyhlml.hlmlShutdown()
    except pyhlml.hlml_error.HLMLError_AlreadyInitialized:
        pass
 
...


        with hlmlinit():
            deviceCount = pyhlml.hlmlDeviceGetCount()
            for i in range(0, deviceCount):
                self.handles[i] = pyhlml.hlmlDeviceGetHandleByIndex(i)

...


    def get_gpus_info(self, selection=None):
        with hlmlinit():
            return make_gpu_infos(self.handles, selection)

        | Traceback (most recent call last):
        |   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 137, in <module>
        |     main()
        |   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 128, in main
        |     trainer.train()
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/ppov2_trainer.py", line 309, in train
        |     query_responses, logitss = batch_generation(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        |     return func(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1176, in batch_generation
        |     query_response, logits = generate(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1151, in generate
        |     output = lm_backbone.generate(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        |     return func(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 2024, in generate
        |     result = self._sample(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 3038, in _sample
        |     unfinished_sequences = unfinished_sequences & ~stopping_criteria(input_ids, scores)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/stopping_criteria.py", line 509, in __call__
        |     is_done = torch.full((input_ids.shape[0],), False, device=input_ids.device, dtype=torch.bool)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/habana_frameworks/torch/gpu_migration/torch/_C/_VariableFunctions.py", line 244, in full
        |     return TorchModuleRegister.full(*args, **kwargs)
        | RuntimeError: [Rank:0] FATAL ERROR :: MODULE:PT_BRIDGE Exception in acc thread pool task has been thrown: Got a non-HPU tensor, expecting an HPU tensor
        | [Rank:0] Habana exception raised from GetHbLazyTensor at aten_lazy_bridge.cpp:290

@Delaunay
Copy link
Collaborator Author

Delaunay commented Oct 9, 2024

Eager Mode

        | Traceback (most recent call last):
        |   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 137, in <module>
        |     main()
        |   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 128, in main
        |     trainer.train()
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/ppov2_trainer.py", line 309, in train
        |     query_responses, logitss = batch_generation(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        |     return func(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1176, in batch_generation
        |     query_response, logits = generate(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1151, in generate
        |     output = lm_backbone.generate(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
        |     return func(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 2024, in generate
        |     result = self._sample(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 2982, in _sample
        |     outputs = self(**model_inputs, return_dict=True)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in _wrapped_call_impl
        |     return self._call_impl(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in _call_impl
        |     return forward_call(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", line 989, in forward
        |     outputs = self.gpt_neox(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in _wrapped_call_impl
        |     return self._call_impl(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in _call_impl
        |     return forward_call(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", line 880, in forward
        |     outputs = layer(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in _wrapped_call_impl
        |     return self._call_impl(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in _call_impl
        |     return forward_call(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", line 648, in forward
        |     self.input_layernorm(hidden_states),
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in _wrapped_call_impl
        |     return self._call_impl(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in _call_impl
        |     return forward_call(*args, **kwargs)
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/normalization.py", line 201, in forward
        |     return F.layer_norm(
        |   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/functional.py", line 2573, in layer_norm
        |     return torch.layer_norm(input, normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled)
        | RuntimeError: Expected all tensors to be on the HPU device, but found at least one input[idx=2] on cpu (details: torch.FloatTensor)

@Delaunay Delaunay added the HPU label Oct 9, 2024
@Delaunay Delaunay changed the title rlhf rlhf HPU Oct 9, 2024
@Delaunay Delaunay changed the title rlhf HPU [HPU] rlhf Oct 15, 2024
@Delaunay
Copy link
Collaborator Author

Delaunay commented Oct 15, 2024 •
edited
Loading 

        import torchcompat.core as comp
        accelerator = comp.accelerate.Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps)
        print(accelerator)

@Delaunay
Copy link
Collaborator Author

Delaunay commented Oct 15, 2024 •
edited
Loading

Eager Mode does not work

rlhf-single.D0 [stderr] Traceback (most recent call last):
rlhf-single.D0 [stderr]   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 137, in <module>
rlhf-single.D0 [stderr]     main()
rlhf-single.D0 [stderr]   File "/homes/delaunap/milabench/benchmarks/rlhf/main.py", line 128, in main
rlhf-single.D0 [stderr]     trainer.train()
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/ppov2_trainer.py", line 323, in train
rlhf-single.D0 [stderr]     query_responses, logitss = batch_generation(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in 
decorate_context
rlhf-single.D0 [stderr]     return func(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1178, in 
batch_generation
rlhf-single.D0 [stderr]     query_response, logits = generate(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/trl/trainer/utils.py", line 1153, in generate
rlhf-single.D0 [stderr]     output = lm_backbone.generate(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in 
decorate_context
rlhf-single.D0 [stderr]     return func(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 1989, in 
generate
rlhf-single.D0 [stderr]     result = self._sample(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/generation/utils.py", line 2932, in 
_sample
rlhf-single.D0 [stderr]     outputs = self(**model_inputs, return_dict=True)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in 
_wrapped_call_impl
rlhf-single.D0 [stderr]     return self._call_impl(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in 
_call_impl
rlhf-single.D0 [stderr]     return forward_call(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", 
line 989, in forward
rlhf-single.D0 [stderr]     outputs = self.gpt_neox(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in 
_wrapped_call_impl
rlhf-single.D0 [stderr]     return self._call_impl(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in 
_call_impl
rlhf-single.D0 [stderr]     return forward_call(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", 
line 880, in forward
rlhf-single.D0 [stderr]     outputs = layer(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in 
_wrapped_call_impl
rlhf-single.D0 [stderr]     return self._call_impl(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in 
_call_impl
rlhf-single.D0 [stderr]     return forward_call(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", 
line 647, in forward
rlhf-single.D0 [stderr]     attention_layer_outputs = self.attention(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1535, in 
_wrapped_call_impl
rlhf-single.D0 [stderr]     return self._call_impl(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1544, in 
_call_impl
rlhf-single.D0 [stderr]     return forward_call(*args, **kwargs)
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", 
line 439, in forward
rlhf-single.D0 [stderr]     query, key, value, present = self._attn_projections_and_rope(
rlhf-single.D0 [stderr]   File "/homes/delaunap/hpu/results/venv/torch/lib/python3.10/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", 
line 222, in _attn_projections_and_rope
rlhf-single.D0 [stderr]     query_pass = query[..., self.rotary_ndims :]
rlhf-single.D0 [stderr] RuntimeError: Graph compile failed. synStatus=synStatus 26 [Generic failure].
  0%|          | 0/254 [00:00<?, ?it/s]
rlhf-single.D0 [end (1)] /homes/delaunap/milabench/benchmarks/rlhf/main.py --output_dir /homes/delaunap/hpu/results/extra/rlhf-single/output 
--model_name_or_path EleutherAI/pythia-1b-deduped --per_device_train_batch_size 64 --logging_strategy no --log_level critical --bf16 [at 2024-10-15

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
HPU
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
1 participant
@Delaunay