improve SQ to adapt optimum-intel 1.16.1 (#1591)

Signed-off-by: changwangss <chang1.wang@intel.com>

Author: changwangss

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

@@ -7,8 +7,7 @@ sentencepiece != 0.1.92
 torch==2.3.0+cpu
 transformers
 intel_extension_for_pytorch==2.3.0
-git+https://github.com/huggingface/optimum.git@e38d40ad220a180213f99b1d93d0407a826c326d
-optimum-intel
+optimum-intel==1.16.1
 bitsandbytes  #baichuan
 transformers_stream_generator
 tiktoken  #qwen

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

@@ -221,6 +221,7 @@
     from intel_extension_for_transformers.transformers.llm.evaluation.models import (
         TSModelCausalLMForITREX,
     )
+
     if args.restore:
         from intel_extension_for_transformers.transformers.utils.utility import (
             recover_model_from_json,
@@ -231,11 +232,13 @@
             args.trust_remote_code,
         )
     else:
-        user_model = TSModelCausalLMForITREX.from_pretrained(
-            args.model,
-            file_name="best_model.pt",
-            trust_remote_code=args.trust_remote_code,
-        )
+        user_model = torch.jit.load(os.path.join( args.model, "best_model.pt"))
+        config = AutoConfig.from_pretrained(args.model, trust_remote_code=args.trust_remote_code)
+        origin_model_type = config.model_type
+        if origin_model_type in ["chatglm", "qwen", "baichuan"]:
+            config.model_type = "qwen2"
+        user_model = TSModelCausalLMForITREX(user_model, config=config)
+        user_model.config.model_type = origin_model_type
 elif not (args.sq or args.mixed_precision):
     user_model = AutoModelForCausalLM.from_pretrained(
         args.model,

intel_extension_for_transformers/transformers/utils/utility.py

@@ -98,52 +98,64 @@ def __init__(self) -> None:
 def generate_dummy_past_key_values(config, input_bs):
     """Generate the dummy past_key_values."""
     from optimum.utils import NormalizedConfigManager
-
-    normalized_config = NormalizedConfigManager.get_normalized_config_class(
-        config.model_type
-    )(config)
-    nb_pkv = 2
-    num_layers = normalized_config.num_layers
-    num_attention_heads = normalized_config.num_attention_heads
-    hidden_size = normalized_config.hidden_size
-    d_k = hidden_size // num_attention_heads
-    num_key_value_heads = num_attention_heads
-    if hasattr(normalized_config, "num_key_value_heads"):
-        num_key_value_heads = normalized_config.num_key_value_heads
-    if hasattr(normalized_config, "multi_query_group_num"):
-        num_key_value_heads = normalized_config.multi_query_group_num
-
-    if config.model_type == "bloom":
-        shape_key = (input_bs * num_attention_heads, d_k, 1)
-        shape_value = (input_bs * num_attention_heads, 1, d_k)
-        key = torch.ones(size=shape_key)
-        value = torch.ones(size=shape_value)
-        past_key_values = tuple(
-            tuple(key if idx % 2 == 0 else value for idx in range(nb_pkv))
-            for _ in range(num_layers)
-        )
-        return past_key_values
-    elif config.model_type == "gpt_bigcode":
-        new_shape = [input_bs, 0, d_k * 2]
-        dummy_tensor = torch.zeros(size=new_shape)
-        past_key_values = tuple([dummy_tensor] * num_layers)
-        return past_key_values
-    elif config.model_type == "qwen":
-        new_shape = [input_bs, 1, num_key_value_heads, d_k]
-        past_key_values = [
-            (
-                torch.ones(size=new_shape).contiguous(),
-                torch.ones(size=new_shape).contiguous(),
-            )
-            for _ in range(num_layers)
+    if config.model_type == "qwen":
+        new_shape = [
+            input_bs,
+            0,
+            config.num_attention_heads,
+            config.hidden_size // config.num_attention_heads,
         ]
-        return tuple(past_key_values)
+        num_layers = config.num_hidden_layers
+    elif config.model_type == "baichuan":
+        new_shape = [
+            input_bs,
+            config.num_attention_heads,
+            0,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_hidden_layers
     elif config.model_type == "chatglm":
-        new_shape = [0, input_bs, num_key_value_heads, d_k]
-    elif config.model_type == "falcon":
-        new_shape = [input_bs, 1, 0, d_k]
+        new_shape = [
+            0,
+            input_bs,
+            config.num_attention_heads,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_layers
     else:
-        new_shape = [input_bs, num_key_value_heads, 0, d_k]
+        normalized_config = NormalizedConfigManager.get_normalized_config_class(
+            config.model_type
+        )(config)
+        nb_pkv = 2
+        num_layers = normalized_config.num_layers
+        num_attention_heads = normalized_config.num_attention_heads
+        hidden_size = normalized_config.hidden_size
+        d_k = hidden_size // num_attention_heads
+        num_key_value_heads = num_attention_heads
+        if hasattr(normalized_config, "num_key_value_heads"):
+            num_key_value_heads = normalized_config.num_key_value_heads
+        if hasattr(normalized_config, "multi_query_group_num"):
+            num_key_value_heads = normalized_config.multi_query_group_num
+
+        if config.model_type == "bloom":
+            shape_key = (input_bs * num_attention_heads, d_k, 1)
+            shape_value = (input_bs * num_attention_heads, 1, d_k)
+            key = torch.ones(size=shape_key)
+            value = torch.ones(size=shape_value)
+            past_key_values = tuple(
+                tuple(key if idx % 2 == 0 else value for idx in range(nb_pkv))
+                for _ in range(num_layers)
+            )
+            return past_key_values
+        elif config.model_type == "gpt_bigcode":
+            new_shape = [input_bs, 0, d_k * 2]
+            dummy_tensor = torch.zeros(size=new_shape)
+            past_key_values = tuple([dummy_tensor] * num_layers)
+            return past_key_values
+        elif config.model_type == "falcon":
+            new_shape = [input_bs, 1, 0, d_k]
+        else:
+            new_shape = [input_bs, num_key_value_heads, 0, d_k]
     past_key_values = [
         (
             torch.zeros(size=new_shape).contiguous(),
@@ -156,44 +168,64 @@ def generate_dummy_past_key_values(config, input_bs):
 def generate_dummy_past_key_values_for_inference(config, input_bs):
     """Generate the dummy past_key_values."""
     from optimum.utils import NormalizedConfigManager
-
-    normalized_config = NormalizedConfigManager.get_normalized_config_class(
-        config.model_type
-    )(config)
-    nb_pkv = 2
-    num_layers = normalized_config.num_layers
-    num_attention_heads = normalized_config.num_attention_heads
-    hidden_size = normalized_config.hidden_size
-    d_k = hidden_size // num_attention_heads
-    num_key_value_heads = num_attention_heads
-    if hasattr(normalized_config, "num_key_value_heads"):
-        num_key_value_heads = normalized_config.num_key_value_heads
-    if hasattr(normalized_config, "multi_query_group_num"):
-        num_key_value_heads = normalized_config.multi_query_group_num
-
-    if config.model_type == "bloom":
-        shape_key = (input_bs * num_attention_heads, d_k, 0)
-        shape_value = (input_bs * num_attention_heads, 0, d_k)
-        key = torch.empty(size=shape_key)
-        value = torch.empty(size=shape_value)
-        past_key_values = tuple(
-            tuple(key if idx % 2 == 0 else value for idx in range(nb_pkv))
-            for _ in range(num_layers)
-        )
-        return past_key_values
-    elif config.model_type == "gpt_bigcode":
-        new_shape = [input_bs, 0, d_k * 2]
-        dummy_tensor = torch.zeros(size=new_shape)
-        past_key_values = tuple([dummy_tensor] * num_layers)
-        return past_key_values
-    elif config.model_type == "qwen":
-        new_shape = [input_bs, 0, num_key_value_heads, d_k]
+    if config.model_type == "qwen":
+        new_shape = [
+            input_bs,
+            0,
+            config.num_attention_heads,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_hidden_layers
+    elif config.model_type == "baichuan":
+        new_shape = [
+            input_bs,
+            config.num_attention_heads,
+            0,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_hidden_layers
     elif config.model_type == "chatglm":
-        new_shape = [0, input_bs, num_key_value_heads, d_k]
-    elif config.model_type == "falcon":
-        new_shape = [input_bs, 1, 0, d_k]
+        new_shape = [
+            0,
+            input_bs,
+            config.num_attention_heads,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_layers
     else:
-        new_shape = [input_bs, num_key_value_heads, 0, d_k]
+        normalized_config = NormalizedConfigManager.get_normalized_config_class(
+            config.model_type
+        )(config)
+        nb_pkv = 2
+        num_layers = normalized_config.num_layers
+        num_attention_heads = normalized_config.num_attention_heads
+        hidden_size = normalized_config.hidden_size
+        d_k = hidden_size // num_attention_heads
+        num_key_value_heads = num_attention_heads
+        if hasattr(normalized_config, "num_key_value_heads"):
+            num_key_value_heads = normalized_config.num_key_value_heads
+        if hasattr(normalized_config, "multi_query_group_num"):
+            num_key_value_heads = normalized_config.multi_query_group_num
+
+        if config.model_type == "bloom":
+            shape_key = (input_bs * num_attention_heads, d_k, 0)
+            shape_value = (input_bs * num_attention_heads, 0, d_k)
+            key = torch.empty(size=shape_key)
+            value = torch.empty(size=shape_value)
+            past_key_values = tuple(
+                tuple(key if idx % 2 == 0 else value for idx in range(nb_pkv))
+                for _ in range(num_layers)
+            )
+            return past_key_values
+        elif config.model_type == "gpt_bigcode":
+            new_shape = [input_bs, 0, d_k * 2]
+            dummy_tensor = torch.zeros(size=new_shape)
+            past_key_values = tuple([dummy_tensor] * num_layers)
+            return past_key_values
+        elif config.model_type == "falcon":
+            new_shape = [input_bs, 1, 0, d_k]
+        else:
+            new_shape = [input_bs, num_key_value_heads, 0, d_k]
     past_key_values = [
         (
             torch.zeros(size=new_shape).contiguous(),
@@ -206,32 +238,53 @@ def generate_dummy_past_key_values_for_inference(config, input_bs):
 def generate_dummy_past_key_values_for_opt_llm(config, input_bs, num_beams=1):
     """Generate the dummy past_key_values."""
     from optimum.utils import NormalizedConfigManager
-
-    normalized_config = NormalizedConfigManager.get_normalized_config_class(
-        config.model_type
-    )(config)
-    num_layers = normalized_config.num_layers
-    num_attention_heads = normalized_config.num_attention_heads
-    hidden_size = normalized_config.hidden_size
-    d_k = hidden_size // num_attention_heads
-    num_key_value_heads = num_attention_heads
-    nb_pkv = 2
-    if hasattr(normalized_config, "num_key_value_heads"):
-        num_key_value_heads = normalized_config.num_key_value_heads
-    if hasattr(normalized_config, "multi_query_group_num"):
-        num_key_value_heads = normalized_config.multi_query_group_num
-    if config.model_type == "bloom":
-        for nb_pkv in range(nb_pkv):
-            if nb_pkv % 2 == 0:
-                new_shape = [input_bs * num_key_value_heads, d_k, 1]
-            else:
-                new_shape = [input_bs * num_key_value_heads, 1, d_k]
-    elif config.model_type == "qwen":
-        new_shape = [input_bs, 1, num_key_value_heads, d_k]
+    if config.model_type == "qwen":
+        new_shape = [
+            input_bs,
+            1,
+            config.num_attention_heads,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_hidden_layers
+    elif config.model_type == "baichuan":
+        new_shape = [
+            input_bs,
+            config.num_attention_heads,
+            1,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_hidden_layers
     elif config.model_type == "chatglm":
-        new_shape = [1, input_bs, num_key_value_heads, d_k]
+        new_shape = [
+            1,
+            input_bs,
+            config.num_attention_heads,
+            config.hidden_size // config.num_attention_heads,
+        ]
+        num_layers = config.num_layers
     else:
-        new_shape = [input_bs, num_key_value_heads, 1, d_k]
+        normalized_config = NormalizedConfigManager.get_normalized_config_class(
+            config.model_type
+        )(config)
+        num_layers = normalized_config.num_layers
+        num_attention_heads = normalized_config.num_attention_heads
+        hidden_size = normalized_config.hidden_size
+        d_k = hidden_size // num_attention_heads
+        num_key_value_heads = num_attention_heads
+        nb_pkv = 2
+        if hasattr(normalized_config, "num_key_value_heads"):
+            num_key_value_heads = normalized_config.num_key_value_heads
+        if hasattr(normalized_config, "multi_query_group_num"):
+            num_key_value_heads = normalized_config.multi_query_group_num
+        if config.model_type == "bloom":
+            for nb_pkv in range(nb_pkv):
+                if nb_pkv % 2 == 0:
+                    new_shape = [input_bs * num_key_value_heads, d_k, 1]
+                else:
+                    new_shape = [input_bs * num_key_value_heads, 1, d_k]
+
+        else:
+            new_shape = [input_bs, num_key_value_heads, 1, d_k]
 
     beam_idx_tmp = torch.zeros(
         (2048, int(input_bs * num_beams)), dtype=torch.long