[LLM] Fix llm models extension issue (#955)

Author: changwangss

examples/huggingface/pytorch/code-generation/quantization/README.md

@@ -83,7 +83,7 @@ python run_generation.py \
     --allow_code_execution \
     --temperature 0.2 \
     --do_sample \
-    --tasks "humaneval" \
+    --tasks "humaneval"
 # mixedprecision
 python run_generation.py \
     --model bigcode/starcoder \
@@ -94,7 +94,7 @@ python run_generation.py \
     --allow_code_execution \
     --temperature 0.2 \
     --do_sample \
-    --tasks "humaneval" \
+    --tasks "humaneval"
 # smoothquant
 # [alternative] --int8 is used for int8 only, --int8_bf16_mixed is used for int8 mixed bfloat16 precision.
 python run_generation.py \
@@ -108,7 +108,7 @@ python run_generation.py \
     --allow_code_execution \
     --temperature 0.2 \
     --do_sample \
-    --tasks "humaneval" \
+    --tasks "humaneval"
 # weightonlyquant
 python run_generation.py \
     --model bigcode/starcoder \
@@ -120,7 +120,7 @@ python run_generation.py \
     --allow_code_execution \
     --temperature 0.2 \
     --do_sample \
-    --tasks "humaneval" \
+    --tasks "humaneval"
 # load_in_4bit
 python run_generation.py \
     --model bigcode/starcoder \
@@ -131,7 +131,7 @@ python run_generation.py \
     --allow_code_execution \
     --temperature 0.2 \
     --do_sample \
-    --tasks "humaneval" \
+    --tasks "humaneval"
 # load_in_8bit
 python run_generation.py \
     --model bigcode/starcoder \
@@ -142,7 +142,7 @@ python run_generation.py \
     --allow_code_execution \
     --temperature 0.2 \
     --do_sample \
-    --tasks "humaneval" \
+    --tasks "humaneval"
 ```
 
 >Note:

examples/huggingface/pytorch/code-generation/quantization/requirements.txt

@@ -7,6 +7,7 @@ sentencepiece != 0.1.92
 torch==2.1.0+cpu
 peft==0.6.2
 transformers >= 4.35.0
+tiktoken #code_gen
 neural-compressor
 intel_extension_for_pytorch
 git+https://github.com/huggingface/optimum.git@927e94739447b13f7eefe085c8d3662654b6a11c

examples/huggingface/pytorch/code-generation/quantization/run_generation.py

@@ -28,7 +28,7 @@
     "--model", nargs="?", default="bigcode/starcoderbase", const="bigcode/starcoderbase"
 )
 parser.add_argument("--trust_remote_code", default=False)
-parser.add_argument("--revision", default="main", type=str)
+parser.add_argument("--_commit_hash", default="main", type=str)
 parser.add_argument("--dataset", nargs="?", default="mbpp", const="mbpp")
 parser.add_argument("--dtype", type=str, default="int8")
 parser.add_argument(
@@ -137,7 +137,9 @@
     args.model,
     truncation_side="left",
     padding_side="right",
+    trust_remote_code=args.trust_remote_code
 )
+
 config = AutoConfig.from_pretrained(
     args.model,
     torchscript=True
@@ -149,7 +151,7 @@
     else False,  # torchscript will force `return_dict=False` to avoid jit errors
     use_cache=True,  # to use kv cache.
     trust_remote_code=args.trust_remote_code,
-    revision=args.revision,
+    _commit_hash=args._commit_hash,
 )
 if not tokenizer.eos_token:
     if tokenizer.bos_token:
@@ -206,7 +208,7 @@
         args.model,
         quantization_config=quantization_config,
         trust_remote_code=args.trust_remote_code,
-        revision=args.revision,
+        _commit_hash=args._commit_hash,
         use_llm_runtime=False,
     )
 elif args.load_in_4bit or args.load_in_8bit:
@@ -215,15 +217,15 @@
         args.model,
         load_in_4bit=args.load_in_4bit,
         load_in_8bit=args.load_in_8bit,
-        revision=args.revision,
+        _commit_hash=args._commit_hash,
         use_llm_runtime=False,
     )
 elif not args.int8 and not args.int8_bf16_mixed:
     user_model = AutoModelForCausalLM.from_pretrained(
         args.model,
         config=config,
         trust_remote_code=args.trust_remote_code,
-        revision=args.revision,
+        _commit_hash=args._commit_hash,
         use_llm_runtime=False,
     )
 
@@ -248,7 +250,7 @@
         args.output_dir,
         file_name="best_model.pt",
         trust_remote_code=args.trust_remote_code,
-        revision=args.revision,
+        _commit_hash=args._commit_hash,
     )
 
 if args.benchmark:

examples/huggingface/pytorch/text-generation/quantization/requirements.txt

@@ -4,9 +4,13 @@ peft
 protobuf
 sentencepiece != 0.1.92
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch==2.1.0+cpu
+torch==2.1.1+cpu
 transformers
 intel_extension_for_pytorch
+bitsandbytes  #baichuan
+transformers_stream_generator
+tiktoken  #qwen
+einops  #qwen
 git+https://github.com/intel/neural-compressor.git
 git+https://github.com/huggingface/optimum-intel.git@f95dea1ae8966dee4d75d622e7b2468c514ba02d
 git+https://github.com/huggingface/optimum.git@927e94739447b13f7eefe085c8d3662654b6a11c

examples/huggingface/pytorch/text-generation/quantization/run_benchmark.sh

@@ -129,7 +129,7 @@ function run_benchmark {
     elif [ "${topology}" = "baichuan_13b" ]; then
         model_name_or_path="baichuan-inc/Baichuan-13B-Base"
         extra_cmd=$extra_cmd" --trust_remote_code True"
-        extra_cmd=$extra_cmd" --revision 14d5b0e204542744900f6fb52422c6d633bdcb00"
+        extra_cmd=$extra_cmd" --_commit_hash 14d5b0e204542744900f6fb52422c6d633bdcb00"
         pip install transformers==4.33
     elif [ "${topology}" = "baichuan2_7b" ]; then
         model_name_or_path="baichuan-inc/Baichuan2-7B-Base"
@@ -142,12 +142,16 @@ function run_benchmark {
     elif [ "${topology}" = "qwen_7b" ]; then
         model_name_or_path="Qwen/Qwen-7B"
         extra_cmd=$extra_cmd" --trust_remote_code True"
+        extra_cmd=$extra_cmd" --_commit_hash f7bc352f27bb1c02ee371a4576942a7d96c8bb97"
+	pip install transformers==4.35.2
     elif [ "${topology}" = "mistral_7b" ]; then
         model_name_or_path="Intel/neural-chat-7b-v3"
     elif [ "${topology}" = "phi_1b" ]; then
         model_name_or_path="susnato/phi-1_dev"
+	pip install transformers==4.36.1
     elif [ "${topology}" = "phi_1_5b" ]; then
         model_name_or_path="susnato/phi-1_5_dev"
+	pip install transformers==4.36.1
     fi
 
     if [[ ${int8} == "true" ]]; then

examples/huggingface/pytorch/text-generation/quantization/run_generation.py

@@ -125,7 +125,7 @@
 # ============AutoModel parameters==============
 parser.add_argument("--load_in_4bit", type=bool, default=False)
 parser.add_argument("--load_in_8bit", type=bool, default=False)
-parser.add_argument("--revision", default="main", type=str)
+parser.add_argument("--_commit_hash", default="main", type=str)
 parser.add_argument("--trust_remote_code", default=False)
 parser.add_argument("--use_llm_runtime", action="store_true")
 # =======================================
@@ -156,7 +156,7 @@
     else False,  # torchscript will force `return_dict=False` to avoid jit errors
     use_cache=True,  # to use kv cache.
     trust_remote_code=args.trust_remote_code,
-    revision=args.revision,
+    _commit_hash=args._commit_hash,
 )
 
 # chatglm
@@ -255,32 +255,33 @@
         args.model,
         quantization_config=quantization_config,
         trust_remote_code=args.trust_remote_code,
-        revision=args.revision,
+        _commit_hash=args._commit_hash,
         use_llm_runtime=args.use_llm_runtime,
+
     )
 elif args.load_in_4bit or args.load_in_8bit:
     # CPU device usage is provided by intel-extension-for-transformers.
     user_model = AutoModelForCausalLM.from_pretrained(
         args.model,
         load_in_4bit=args.load_in_4bit,
         load_in_8bit=args.load_in_8bit,
-        revision=args.revision,
+        _commit_hash=args._commit_hash,
         use_llm_runtime=args.use_llm_runtime,
     )
 elif (not args.int8 and not args.int8_bf16_mixed) or args.restore:
     if args.peft_model_id is not None:
         user_model = AutoModelForCausalLM.from_pretrained(
             args.peft_model_id,
             trust_remote_code=args.trust_remote_code,
-            revision=args.revision,
+            _commit_hash=args._commit_hash,
             use_llm_runtime=args.use_llm_runtime,
         )
     else:
         user_model = AutoModelForCausalLM.from_pretrained(
             args.model,
             config=config,
             trust_remote_code=args.trust_remote_code,
-            revision=args.revision,
+            _commit_hash=args._commit_hash,
             use_llm_runtime=args.use_llm_runtime,
         )
 
@@ -389,8 +390,8 @@
         + ",tokenizer="
         + args.model
         + ",dtype=float32"
-        + ",revision="
-        + args.revision
+        + ",_commit_hash="
+        + args._commit_hash
         + ",trust_remote_code="
         + str(args.trust_remote_code),
         user_model=user_model,

examples/huggingface/pytorch/text-generation/quantization/run_tuning.sh

@@ -145,7 +145,7 @@ function run_tuning {
         model_name_or_path="tiiuae/falcon-7b-instruct"
         extra_cmd=$extra_cmd" --sq --alpha ${alpha}"
         extra_cmd=$extra_cmd" --output_dir ${tuned_checkpoint}"
-        pip install transformers==4.33
+	pip install transformers==4.33.3
     elif [ "${topology}" = "baichuan_7b" ]; then
         alpha=0.85
         model_name_or_path="baichuan-inc/Baichuan-7B"
@@ -159,7 +159,7 @@ function run_tuning {
         extra_cmd=$extra_cmd" --sq --alpha ${alpha}"
         extra_cmd=$extra_cmd" --output_dir ${tuned_checkpoint}"
         extra_cmd=$extra_cmd" --trust_remote_code True"
-        extra_cmd=$extra_cmd" --revision 14d5b0e204542744900f6fb52422c6d633bdcb00"
+        extra_cmd=$extra_cmd" --_commit_hash 14d5b0e204542744900f6fb52422c6d633bdcb00"
         pip install transformers==4.33
     elif [ "${topology}" = "baichuan2_7b" ]; then
         alpha=0.85
@@ -181,6 +181,8 @@ function run_tuning {
         extra_cmd=$extra_cmd" --sq --alpha ${alpha}"
         extra_cmd=$extra_cmd" --output_dir ${tuned_checkpoint}"
         extra_cmd=$extra_cmd" --trust_remote_code True"
+        extra_cmd=$extra_cmd" --_commit_hash f7bc352f27bb1c02ee371a4576942a7d96c8bb97"
+	pip install transformers==4.35.2
     elif [ "${topology}" = "mistral_7b" ]; then
         alpha=0.8
         model_name_or_path="Intel/neural-chat-7b-v3"
@@ -193,12 +195,14 @@ function run_tuning {
         extra_cmd=$extra_cmd" --sq --alpha ${alpha}"
         extra_cmd=$extra_cmd" --output_dir ${tuned_checkpoint}"
         extra_cmd=$extra_cmd" --trust_remote_code True"
+	pip install transformers==4.36.1
     elif [ "${topology}" = "phi_1_5b" ]; then
         alpha=0.5
         model_name_or_path="susnato/phi-1_5_dev"
         extra_cmd=$extra_cmd" --sq --alpha ${alpha}"
         extra_cmd=$extra_cmd" --output_dir ${tuned_checkpoint}"
         extra_cmd=$extra_cmd" --trust_remote_code True"
+	pip install tranformers==4.36.1
     fi
 
     if [ ${script} = "run_generation.py" ];then

intel_extension_for_transformers/llm/evaluation/lm_eval/models/huggingface.py

@@ -115,7 +115,8 @@ def __init__(
         bnb_4bit_compute_dtype: Optional[Union[str, torch.dtype]] = None,
         bnb_4bit_use_double_quant: Optional[bool] = False,
         init_empty_weights: Optional[bool] = False,
-        model_format: Optional[str] = "torch"
+        model_format: Optional[str] = "torch",
+        _commit_hash: Optional[str] = None
     ):
         """Initializes a HuggingFace `AutoModel` and `AutoTokenizer` for evaluation.
         Args:

intel_extension_for_transformers/llm/evaluation/models.py

@@ -22,7 +22,7 @@
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from optimum.intel.generation.modeling import TSModelForCausalLM
 from intel_extension_for_transformers.transformers.utils.utility import (
-    generate_dummy_past_key_values,
+    generate_dummy_past_key_values_for_inference,
     generate_dummy_past_key_values_for_opt_llm,
     MODEL_TYPES_REQUIRING_POSITION_IDS,
     IPEX_OPT_LLM_SUPPORTED,
@@ -166,18 +166,16 @@ def forward(
         input_bs, input_len = input_ids.shape
         if self.use_cache and past_key_values is None:
             if model_type in IPEX_OPT_LLM_SUPPORTED:
-                if (model_type == "falcon" and transformers.__version__ > "4.33") or (
-                    model_type == "llama" and transformers.__version__ >= "4.36"
-                ):
-                    past_key_values = generate_dummy_past_key_values(
+                if model_type == "llama" and transformers.__version__ >= "4.36":
+                    past_key_values = generate_dummy_past_key_values_for_inference(
                         config=self.config, input_bs=input_bs
                     )
                 else:
                     past_key_values = generate_dummy_past_key_values_for_opt_llm(
                         config=self.config, input_bs=input_bs, num_beams=1
                     )
             else:
-                past_key_values = generate_dummy_past_key_values(
+                past_key_values = generate_dummy_past_key_values_for_inference(
                     config=self.config, input_bs=input_bs
                 )
         inputs["past_key_values"] = past_key_values
@@ -195,7 +193,6 @@ def forward(
                 inputs["position_ids"] = position_ids
             else:
                 inputs["position_ids"] = torch.arange(input_len).repeat(input_bs, 1)
-
         outputs = self.model(**inputs)
 
         if isinstance(outputs, (list, tuple)):

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -228,11 +228,10 @@ def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
                 model = model.float()
             model.eval()
             model_type = model.config.model_type.replace("_", "-")
-            if "falcon" in model_type and transformers.__version__ > "4.33":
-                ipex.nn.utils._model_convert.replace_customized_linear_with_linear(
-                    model.eval()
+            if "falcon" in model_type:
+                logger.warning(
+                    "Please use transformers 4.33.3 if you would like to apply smoothquant to Falcon."
                 )
-                quantization_config.ipex_opt_llm = False
             if "llama" in model_type and transformers.__version__ >= "4.36.0":
                 quantization_config.ipex_opt_llm = False
             logger.info("Applying SmoothQuant.")
@@ -334,7 +333,11 @@ def collate_batch(batch):
                             )
 
                         last_ind.append(input_ids.shape[0] - 1)
-                        attention_mask = torch.ones(len(input_ids))
+                        if model_type in ["bloom", "qwen"]:
+                            attention_mask = torch.ones(len(input_ids) +1)
+                            attention_mask[0] = 0
+                        else:
+                            attention_mask = torch.ones(len(input_ids))
                         position_ids = torch.arange(len(input_ids))
                         input_ids_padded.append(input_ids)
                         attention_mask_padded.append(attention_mask)
@@ -450,17 +453,6 @@ def calib_func(model):
                                 "position_ids": inputs["position_ids"],
                                 "past_key_values": inputs["past_key_values"],
                             }
-                        elif model_type == "falcon":
-                            input_bs, input_len = inputs["input_ids"].shape
-                            outputs = model(inputs["input_ids"])
-                            example_inputs["past_key_values"] = outputs[1]
-                            example_inputs["attention_mask"] = torch.ones(
-                                input_bs, input_len
-                            )
-                            example_inputs["position_ids"] = (
-                                inputs["position_ids"][:, -1:] + 1
-                            )
-                            example_inputs["input_ids"] = inputs["input_ids"][:, -1:]
                         else:
                             example_inputs = inputs
                     else: