improve HPU usage (#1643)

 refine example
 add hpu in auto accelerator and fix bug
---------

Signed-off-by: xin3he <[email protected]>

Author: xin3he

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/habana_fp8/requirement.txt

@@ -1,5 +1,7 @@
 transformers
 datasets
+accelerate
 SentencePiece
-intel_extension_for_transformers
-lm_eval
+lm_eval==0.3.0
+openpyxl
+einops

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/habana_fp8/run_llm.py

@@ -17,21 +17,15 @@
 import transformers
 from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
 import habana_frameworks.torch.core as htcore
-import numpy as np
-import lm_eval
-import lm_eval.tasks
-import lm_eval.evaluator
 from accelerate import init_empty_weights
-from utils import itrex_bootstrap_stderr, show_msg, save_to_excel
+from utils import show_msg, eval_func
 
 
 torch.set_grad_enabled(False)
 htcore.hpu_set_env()
 torch.device('hpu')
 
 
-# to avoid out-of-memory caused by Popen for large language models.
-lm_eval.metrics.bootstrap_stderr = itrex_bootstrap_stderr
 
 
 parser = argparse.ArgumentParser()
@@ -52,6 +46,7 @@
 parser.add_argument("--precision", type=str, default='fp8_e4m3',
                     help="Select from ['fp8_e4m3', 'fp8_e5m2', 'bf16', 'fp16'], \
                         ['bf16', 'fp16'] only work with cast approach")
+parser.add_argument("--autotune", action="store_true")
 parser.add_argument("--accuracy", action="store_true")
 parser.add_argument("--performance", action="store_true")
 parser.add_argument("--generate", action="store_true")
@@ -182,8 +177,9 @@
 ### dynamic & static quantization ###
 if args.approach in ["dynamic", "static"] and not args.load:
     print("device:", next(user_model.parameters()).device)
-    from neural_compressor.torch.quantization.config import FP8Config, get_default_fp8_config
-    from neural_compressor.torch.quantization import quantize
+    from neural_compressor.torch.quantization import (
+        quantize, autotune, FP8Config, get_default_fp8_config, TuningConfig, get_default_fp8_config_set
+    )
     dtype = args.precision
     if args.approach == "dynamic":
         from neural_compressor.torch.algorithms.habana_fp8 import quantize_dynamic
@@ -300,106 +296,7 @@ def replace_torch_mm_bmm():
 
 
 if args.accuracy:
-
-    class HabanaModelAdapter(lm_eval.base.BaseLM):
-        def __init__(self, tokenizer, model, args, options):
-            super().__init__()
-            self.tokenizer = tokenizer
-            self.model = model.eval()
-            self._batch_size = args.batch_size
-            self.buckets = list(sorted(args.buckets))
-            self.options = options
-            self._device = "hpu"
-            torch.set_grad_enabled(False)
-
-        @property
-        def eot_token_id(self):
-            return self.model.config.eos_token_id
-
-        @property
-        def max_length(self):
-            return self.buckets[-1]
-
-        @property
-        def max_gen_toks(self):
-            raise NotImplementedError()
-
-        @property
-        def batch_size(self):
-            return self._batch_size
-
-        @property
-        def device(self):
-            # We need to do padding ourselves, otherwise we'll end up with recompilations
-            # Returning 'cpu' to keep tensors on CPU in lm_eval code
-            return 'cpu' # 'hpu'
-
-        def tok_encode(self, string):
-            if re.search("chatglm3", args.model.lower()) or re.search("llama", args.model.lower()) :
-                string = string.lstrip()
-            return self.tokenizer.encode(string, add_special_tokens=False)
-
-        def tok_decode(self, tokens):
-            return self.tokenizer.decode(tokens, skip_special_tokens=True)
-
-        def _model_generate(self, context, max_length, eos_token_id):
-            raise NotImplementedError()
-
-        def find_bucket(self, length):
-            return [b for b in self.buckets if b >= length][0]
-
-        def _model_call(self, inps):
-            seq_length = inps.shape[-1]
-            padding_length = 0
-            bucket_length = self.find_bucket(seq_length)
-            padding_length = bucket_length - seq_length
-            inps = F.pad(inps, (0, padding_length), value=self.model.config.pad_token_id)
-            logits = self.model(inps.to(self._device))["logits"].cpu()
-
-            if padding_length > 0:
-                logits = logits[:, :-padding_length, :]
-            logits = logits.to(torch.float32)
-            return logits
-
-    lm_tasks = lm_eval.tasks.get_task_dict(args.tasks)
-    options = None
-    lm = HabanaModelAdapter(tokenizer, user_model, args, options)
-
-    eval_start = time.perf_counter()
-    if args.approach == "cast":
-        from neural_compressor.torch.amp import autocast
-        if args.precision == "fp8_e4m3":
-            dtype = torch.float8_e4m3fn
-        elif args.precision == "fp8_e5m2":
-            dtype = torch.float8_e5m2
-        elif args.precision == "fp16":
-            dtype = torch.float16
-        elif args.precision == "bf16":
-            dtype = torch.bfloat16
-        with autocast('hpu', dtype=dtype):
-            results = lm_eval.evaluator.evaluate(lm, lm_tasks, limit=args.limit)
-    else:
-        results = lm_eval.evaluator.evaluate(lm, lm_tasks, limit=args.limit)
-    print(lm_eval.evaluator.make_table(results))
-    eval_end = time.perf_counter()
-    print("Duration:", eval_end - eval_start)
-    results['args'] = vars(args)
-    results['duration'] = eval_end - eval_start
-
-
-    dumped = json.dumps(results, indent=2)
-    accu_dict = {}
-    case_name =  args.approach + "-" + args.precision
-    for task_name in args.tasks:
-        if task_name == "wikitext":
-            print("Accuracy for %s is: %s" % (task_name, results["results"][task_name]["word_perplexity"]), flush=True)
-            accu_dict[task_name] = [args.model, case_name, results["results"][task_name]["word_perplexity"]]
-        else:
-            print("Accuracy for %s is: %s" % (task_name, results["results"][task_name]["acc"]), flush=True)
-            accu_dict[task_name] = [args.model, case_name, results["results"][task_name]["acc"]]
-    if args.dump_to_excel and local_rank in [-1, 0]:
-        save_to_excel(accu_dict)
-
+    eval_func(user_model, tokenizer=tokenizer, args=args)
 
 # dump final message of HPU
 show_msg()

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/habana_fp8/utils.py

@@ -33,3 +33,125 @@ def save_to_excel(dict):
         df_existing = pd.DataFrame()
     df_combined = pd.concat([df_existing, df_new], axis=0, ignore_index=True)
     df_combined.to_excel('output.xlsx', index=False, engine='openpyxl', header=True)
+
+
+def eval_func(user_model, tokenizer, args):
+    import os
+    import re
+    import time
+    import json
+    import torch
+    import habana_frameworks.torch.hpex
+    import torch.nn.functional as F
+    import lm_eval
+    import lm_eval.tasks
+    import lm_eval.evaluator
+
+    # to avoid out-of-memory caused by Popen for large language models.
+    lm_eval.metrics.bootstrap_stderr = itrex_bootstrap_stderr
+
+    class HabanaModelAdapter(lm_eval.base.BaseLM):
+        def __init__(self, tokenizer, model, args, options):
+            super().__init__()
+            self.tokenizer = tokenizer
+            self.model = model.eval()
+            self._batch_size = args.batch_size
+            self.buckets = list(sorted(args.buckets))
+            self.options = options
+            self._device = "hpu"
+            torch.set_grad_enabled(False)
+
+        @property
+        def eot_token_id(self):
+            return self.model.config.eos_token_id
+
+        @property
+        def max_length(self):
+            return self.buckets[-1]
+
+        @property
+        def max_gen_toks(self):
+            raise NotImplementedError()
+
+        @property
+        def batch_size(self):
+            return self._batch_size
+
+        @property
+        def device(self):
+            # We need to do padding ourselves, otherwise we'll end up with recompilations
+            # Returning 'cpu' to keep tensors on CPU in lm_eval code
+            return 'cpu' # 'hpu'
+
+        def tok_encode(self, string):
+            if (
+                re.search("chatglm3", args.model.lower()) or
+                re.search("llama", args.model.lower()) or
+                re.search("mistral", args.model.lower())
+            ):
+                string = string.lstrip()
+            return self.tokenizer.encode(string, add_special_tokens=False)
+
+        def tok_decode(self, tokens):
+            return self.tokenizer.decode(tokens, skip_special_tokens=True)
+
+        def _model_generate(self, context, max_length, eos_token_id):
+            raise NotImplementedError()
+
+        def find_bucket(self, length):
+            return [b for b in self.buckets if b >= length][0]
+
+        def _model_call(self, inputs):
+            seq_length = inputs.shape[-1]
+            padding_length = 0
+            bucket_length = self.find_bucket(seq_length)
+            padding_length = bucket_length - seq_length
+            inputs = F.pad(inputs, (0, padding_length), value=self.model.config.pad_token_id)
+            logits = self.model(inputs.to(self._device))["logits"].cpu()
+
+            if padding_length > 0:
+                logits = logits[:, :-padding_length, :]
+            logits = logits.to(torch.float32)
+            return logits
+
+    lm_tasks = lm_eval.tasks.get_task_dict(args.tasks)
+    options = None
+    lm = HabanaModelAdapter(tokenizer, user_model, args, options)
+
+    eval_start = time.perf_counter()
+    if args.approach == "cast":
+        from neural_compressor.torch.amp import autocast
+        if args.precision == "fp8_e4m3":
+            dtype = torch.float8_e4m3fn
+        elif args.precision == "fp8_e5m2":
+            dtype = torch.float8_e5m2
+        elif args.precision == "fp16":
+            dtype = torch.float16
+        elif args.precision == "bf16":
+            dtype = torch.bfloat16
+        with autocast('hpu', dtype=dtype):
+            results = lm_eval.evaluator.evaluate(lm, lm_tasks, limit=args.limit)
+    else:
+        results = lm_eval.evaluator.evaluate(lm, lm_tasks, limit=args.limit)
+    print(lm_eval.evaluator.make_table(results))
+    eval_end = time.perf_counter()
+    print("Duration:", eval_end - eval_start)
+    results['args'] = vars(args)
+    results['duration'] = eval_end - eval_start
+
+    # make sure that result is dumped only once during multi-cards evaluation
+    local_rank = int(os.getenv('LOCAL_RANK', '-1'))
+    if local_rank in [-1, 0]:
+        dumped = json.dumps(results, indent=2)
+        accu_dict = {}
+        case_name =  args.approach + "-" + args.precision
+        for task_name in args.tasks:
+            if task_name == "wikitext":
+                print("Accuracy for %s is: %s" % (task_name, results["results"][task_name]["word_perplexity"]), flush=True)
+                accu_dict[task_name] = [args.model, case_name, results["results"][task_name]["word_perplexity"]]
+            else:
+                print("Accuracy for %s is: %s" % (task_name, results["results"][task_name]["acc"]), flush=True)
+                accu_dict[task_name] = [args.model, case_name, results["results"][task_name]["acc"]]
+        if args.dump_to_excel:
+            save_to_excel(accu_dict)
+    return results["results"][task_name]["acc"]

neural_compressor/torch/algorithms/weight_only/gptq.py

@@ -257,8 +257,7 @@ def __init__(
 
         # device
         self.device = get_device(kwargs.pop("device", "auto"))
-        if str(self.model.device).startswith("cuda"):
-            self.device = self.model.device
+        self.model.to(self.device)
         self.is_ready = False
 
         self.export_compressed_model = export_compressed_model

neural_compressor/torch/utils/auto_accelerator.py

@@ -31,7 +31,8 @@
 
 from neural_compressor.torch.utils import logger
 
-PRIORITY_CUDA = 100
+PRIORITY_HPU = 100
+PRIORITY_CUDA = 95
 PRIORITY_CPU = 90
 
 
@@ -53,8 +54,9 @@ class CPU_Accelerator:
         """
 
         def decorator(accelerator_cls):
-            cls.registered_accelerators.setdefault(name, {})
-            cls.registered_accelerators[name] = (accelerator_cls, priority)
+            if accelerator_cls.is_available():
+                cls.registered_accelerators.setdefault(name, {})
+                cls.registered_accelerators[name] = (accelerator_cls, priority)
             return accelerator_cls
 
         return decorator
@@ -202,6 +204,47 @@ def empty_cache(self):
         return torch.cuda.empty_cache()
 
 
+@register_accelerator(name="hpu", priority=PRIORITY_HPU)
+class HPU_Accelerator(Auto_Accelerator):
+    def __init__(self) -> None:
+        self._name = "hpu"
+
+    def name(self) -> str:
+        return self._name
+
+    @classmethod
+    def is_available(cls) -> bool:
+        from .environ import is_hpex_available
+
+        if is_hpex_available():
+            return torch.hpu.is_available()
+        else:
+            return False
+
+    def device_name(self, device_indx) -> str:
+        if device_indx is None:
+            return "hpu"
+        return f"hpu:{device_indx}"
+
+    def synchronize(self):
+        return torch.hpu.synchronize()
+
+    def set_device(self, device_index):
+        return torch.hpu.set_device(device_index)
+
+    def current_device(self):
+        return torch.hpu.current_device()
+
+    def current_device_name(self):
+        return "hpu:{}".format(torch.hpu.current_device())
+
+    def device(self, device_index=None):
+        return torch.hpu.device(device_index)
+
+    def empty_cache(self):
+        return torch.hpu.empty_cache()
+
+
 def auto_detect_accelerator(device_name="auto") -> Auto_Accelerator:
     # Force use the cpu on node has both cpu and gpu: `FORCE_DEVICE=cpu` python main.py ...
     # The `FORCE_DEVICE` is case insensitive.