Merge branch 'master' into xuehao/update_ci

Signed-off-by: Sun, Xuehao <[email protected]>

.azure-pipelines/ut-3x-pt-fp8.yml

@@ -15,6 +15,8 @@ pr:
       - neural_compressor/torch
       - test/3x/torch/algorithms/fp8_quant
       - test/3x/torch/quantization/fp8_quant
+      - test/3x/torch/quantization/weight_only/test_rtn.py
+      - test/3x/torch/quantization/weight_only/test_load.py
       - setup.py
       - requirements_pt.txt
 

README.md

@@ -78,7 +78,7 @@ Following example code demonstrates FP8 Quantization, it is supported by Intel G
 To try on Intel Gaudi2, docker image with Gaudi Software Stack is recommended, please refer to following script for environment setup. More details can be found in [Gaudi Guide](https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#launch-docker-image-that-was-built).
 ```bash
 # Run a container with an interactive shell
-docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.19.0/ubuntu24.04/habanalabs/pytorch-installer-2.5.1:latest
 ```
 Run the example:
 ```python

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/weight_only/requirements-autoround-hpu.txt

@@ -11,6 +11,5 @@ lm_eval==0.4.3
 peft
 numba
 tbb
-# TODO: (Yi) SW-208079 replace auto-round with the released version
-auto-round-hpu @ git+https://github.com/intel/auto-round.git@hpu_only_pkg
-optimum-habana==1.14.1
+auto-round @ git+https://github.com/intel/[email protected]
+optimum-habana==1.14.1

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/weight_only/run_benchmark.sh

@@ -14,6 +14,7 @@ function init_params {
   batch_size=16
   tuned_checkpoint=saved_results
   task=lambada_openai
+  incbench_cmd="incbench --num_cores_per_instance 4"
   echo ${max_eval_samples}
   for var in "$@"
   do
@@ -104,6 +105,7 @@ function run_benchmark {
     elif [ "${topology}" = "opt_125m_woq_autoround_int4_hpu" ]; then
         model_name_or_path="facebook/opt-125m"
         extra_cmd=$extra_cmd" --woq_algo AutoRound"
+        incbench_cmd="incbench --num_instances 1"
     elif [ "${topology}" = "opt_125m_woq_autotune_int4" ]; then
         model_name_or_path="facebook/opt-125m"
     fi
@@ -116,7 +118,7 @@ function run_benchmark {
             --batch_size ${batch_size} \
             ${extra_cmd} ${mode_cmd}
     elif [[ ${mode} == "performance" ]]; then
-        incbench --num_cores_per_instance 4 run_clm_no_trainer.py \
+        ${incbench_cmd} run_clm_no_trainer.py \
             --model ${model_name_or_path} \
             --batch_size ${batch_size} \
             --output_dir ${tuned_checkpoint} \

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/weight_only/run_clm_no_trainer.py

@@ -270,8 +270,9 @@ def get_user_model():
         torchscript = True
     if args.woq_algo == "AutoRound" and is_habana_framework_installed():
         print("Quantizing model with AutoRound on HPU")
-        check_torch_compile_with_hpu_backend()
-        set_envs_for_torch_compile_with_hpu_backend()
+        if args.quantize:
+            check_torch_compile_with_hpu_backend()
+            set_envs_for_torch_compile_with_hpu_backend()
         user_model = AutoModelForCausalLM.from_pretrained(
             args.model,
             trust_remote_code=args.trust_remote_code,
@@ -403,11 +404,12 @@ def calib_func(prepared_model):
             max_seq_length=args.gptq_max_seq_length,
         )
         dataloader_for_calibration = dataloaderPreprocessor.get_prepared_dataloader()
-        from neural_compressor.torch.algorithms.weight_only.utility import move_input_to_device
+        from neural_compressor.torch.utils import get_model_device, move_input_device
         from tqdm import tqdm
         def run_fn_for_gptq(model, dataloader_for_calibration, *args):
             for batch in tqdm(dataloader_for_calibration):
-                batch = move_input_to_device(batch, device=None)
+                device = get_model_device(model)
+                batch = move_input_device(batch, device=device)
                 if isinstance(batch, tuple) or isinstance(batch, list):
                     model(batch[0])
                 elif isinstance(batch, dict):
@@ -525,11 +527,12 @@ def run_fn_for_autoround(model, dataloader):
         )
         dataloader = dataloaderPreprocessor.get_prepared_dataloader()
         custom_tune_config = TuningConfig(config_set=get_woq_tuning_config())
-        from neural_compressor.torch.algorithms.weight_only.utility import move_input_to_device
+        from neural_compressor.torch.utils import get_model_device, move_input_device
         from tqdm import tqdm
         def run_fn_for_gptq(model, dataloader_for_calibration, *args):
             for batch in tqdm(dataloader_for_calibration):
-                batch = move_input_to_device(batch, device=None)
+                device = get_model_device(model)
+                batch = move_input_device(batch, device=device)
                 if isinstance(batch, tuple) or isinstance(batch, list):
                     model(batch[0])
                 elif isinstance(batch, dict):
@@ -568,7 +571,7 @@ def run_fn_for_gptq(model, dataloader_for_calibration, *args):
 
 
 if is_hpex_available():
-    from habana_frameworks.torch.hpu import wrap_in_hpu_graph
+    from habana_frameworks.torch.hpu.graphs import wrap_in_hpu_graph
     user_model = user_model.to(torch.bfloat16)
     wrap_in_hpu_graph(user_model, max_graphs=10)
 

neural_compressor/common/__init__.py

@@ -27,7 +27,7 @@
     dump_elapsed_time,
 )
 from neural_compressor.common.base_config import options
-
+from neural_compressor.common.version import __version__
 
 __all__ = [
     "options",

neural_compressor/common/version.py

@@ -0,0 +1 @@
+../version.py

neural_compressor/evaluation/lm_eval/models/huggingface.py

@@ -969,6 +969,8 @@ def _model_call(self, inps, attn_mask=None, labels=None):
                         output = output.logits
                     if self.pad_to_buckets and padding_length != 0:  # use buckets to pad inputs
                         output = output[:, :-padding_length, :]
+                    if "hpu" in output.device.type:  # make sure return fp32 tensor for HPU, TODO: root cause
+                        output = output.to(torch.float32)
                 return output
 
     def _model_generate(self, context, max_length, stop, **generation_kwargs):

neural_compressor/torch/algorithms/weight_only/save_load.py

@@ -576,7 +576,7 @@ def _get_loaded_state_dict(self, config):
             "_raise_exceptions_for_missing_entries": False,
             "_commit_hash": commit_hash,
         }
-        resolved_archive_file = self._get_resolved_archive_file(**kwargs)
+        resolved_archive_file, is_sharded = self._get_resolved_archive_file(**kwargs)
 
         self._model_local_dir = os.path.abspath(os.path.expanduser(os.path.dirname(resolved_archive_file)))
         # if hpu format tensor can be used directly, then update resolved_archive_file to the hpu format tensor file
@@ -640,6 +640,7 @@ def _get_resolved_archive_file(self, **kwargs):
         subfolder = kwargs.get("subfolder")
 
         resolved_archive_file = None
+        is_sharded = False
         is_local = os.path.isdir(self.model_name_or_path)
         if is_local:  # pragma: no cover
             # self.model_name_or_path is a local directory
@@ -787,7 +788,7 @@ def _get_resolved_archive_file(self, **kwargs):
         if is_local:
             resolved_archive_file = archive_file
 
-        return resolved_archive_file
+        return resolved_archive_file, is_sharded
 
     def _init_hf_model(self, model_class, config):
         from accelerate.big_modeling import init_empty_weights

test/3x/torch/algorithms/fp8_quant/unit_tests/test_functions/test_config_json.py

@@ -1,7 +1,7 @@
 """Use this module as an example of how to write new unit tests for layers."""
 
 import os
-
+import sys
 import pytest
 import torch
 
@@ -58,7 +58,12 @@ def run_predefined_config():
             run_with_raised_exception(run_predefined_config, FileNotFoundError, "Failed to load file ")
         # TODO [SW-196641]: fix the following issue:
         elif quant_mode == QuantMode.SHAPE:
-            run_with_raised_exception(run_predefined_config, UnboundLocalError, "local variable 'fname_base' referenced before assignment")
+            error_message = (
+                "cannot access local variable 'fname_base' where it is not associated with a value"
+                if sys.version_info >= (3, 11)
+                else "local variable 'fname_base' referenced before assignment"
+            )
+            run_with_raised_exception(run_predefined_config, UnboundLocalError, error_message)
     else:
         run_predefined_config()
 

test/3x/torch/quantization/weight_only/test_rtn.py

@@ -16,6 +16,7 @@
 from neural_compressor.torch.utils import accelerator, is_hpex_available
 
 device = accelerator.current_device_name()
+torch.set_grad_enabled(False)
 
 
 class ModelConv1d(torch.nn.Module):

test/3x/torch/requirements.txt

@@ -1,5 +1,5 @@
 auto_round
-deepspeed @ git+https://github.com/HabanaAI/DeepSpeed.git@1.18.0
+deepspeed @ git+https://github.com/HabanaAI/DeepSpeed.git@1.19.0
 expecttest
 intel_extension_for_pytorch
 numpy