Added GPTQ/AWQ support with HF Transformers (#1933)

Author: AlexKoff88

tools/who_what_benchmark/requirements.txt

@@ -8,3 +8,5 @@ numpy>=1.23.5
 tqdm>=4.66.1
 diffusers
 datasets<3.2.0
+auto-gptq; sys_platform != "darwin"
+autoawq<0.2.8; sys_platform != "darwin"

tools/who_what_benchmark/tests/test_cli_text.py

@@ -6,6 +6,7 @@
 import pytest
 import logging
 import json
+import sys
 
 from transformers import AutoTokenizer
 from optimum.intel.openvino import OVModelForCausalLM, OVWeightQuantizationConfig
@@ -27,6 +28,9 @@ def run_wwb(args):
 base_model_path = os.path.join(tmp_dir, "opt125m")
 target_model_path = os.path.join(tmp_dir, "opt125m_int8")
 
+gptq_model_id = "ybelkada/opt-125m-gptq-4bit"
+awq_model_id = "TitanML/tiny-mixtral-AWQ-4bit"
+
 
 def setup_module():
     from optimum.exporters.openvino.convert import export_tokenizer
@@ -181,7 +185,21 @@ def test_text_language():
     assert "马克" in data["prompts"].values[0]
 
 
-def test_text_hf_model():
+hf_model_scope = [
+    (model_id),
+]
+if sys.platform != 'darwin':
+    hf_model_scope += [
+        (gptq_model_id),
+        (awq_model_id),
+    ]
+
+
+@pytest.mark.parametrize(
+    ("model_id"),
+    hf_model_scope,
+)
+def test_text_hf_model(model_id):
     with tempfile.TemporaryDirectory() as temp_dir:
         temp_file_name = os.path.join(temp_dir, "gt.csv")
         result = run_wwb(
@@ -191,15 +209,15 @@ def test_text_hf_model():
                 "--gt-data",
                 temp_file_name,
                 "--num-samples",
-                "2",
+                "1",
                 "--device",
                 "CPU",
                 "--hf",
             ]
         )
         assert result.returncode == 0
         data = pd.read_csv(temp_file_name)
-    assert len(data["prompts"].values) == 2
+    assert len(data["prompts"].values) == 1
 
 
 def test_text_genai_model():

tools/who_what_benchmark/whowhatbench/model_loaders.py

@@ -1,9 +1,12 @@
 import logging
 import json
+import torch
 
 from transformers import AutoConfig, AutoModelForCausalLM, AutoModel, AutoModelForVision2Seq, AutoTokenizer
 from diffusers import DiffusionPipeline, AutoPipelineForImage2Image, AutoPipelineForInpainting
 
+from .utils import mock_torch_cuda_is_available
+
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -82,18 +85,39 @@ def load_text_llamacpp_pipeline(model_dir):
     return model
 
 
+def load_text_hf_pipeline(model_id, device):
+    model_kwargs = {}
+
+    if not torch.cuda.is_available or device.lower() == "cpu":
+        config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
+        is_gptq = False
+        is_awq = False
+        if getattr(config, "quantization_config", None):
+            is_gptq = config.quantization_config["quant_method"] == "gptq"
+            is_awq = config.quantization_config["quant_method"] == "awq"
+        if is_gptq or is_awq:
+            # infer in FP32
+            model_kwargs["torch_dtype"] = torch.float32
+        with mock_torch_cuda_is_available(is_gptq or is_awq):
+            model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, device_map="cpu", **model_kwargs)
+        if is_awq:
+            model.is_awq = is_awq
+    else:
+        model = AutoModelForCausalLM.from_pretrained(
+            model_id, trust_remote_code=True, device_map=device.lower(), **model_kwargs
+        )
+    model.eval()
+    return model
+
+
 def load_text_model(
     model_id, device="CPU", ov_config=None, use_hf=False, use_genai=False, use_llamacpp=False, **kwargs,
 ):
     if use_hf:
         logger.info("Using HF Transformers API")
-        model = AutoModelForCausalLM.from_pretrained(
-            model_id, trust_remote_code=True, device_map=device.lower()
-        )
-        model.eval()
+        model = load_text_hf_pipeline(model_id, device)
     elif use_genai:
         model = load_text_genai_pipeline(model_id, device, ov_config, **kwargs)
-
     elif use_llamacpp:
         logger.info("Using llama.cpp API")
         model = load_text_llamacpp_pipeline(model_id)

tools/who_what_benchmark/whowhatbench/text_evaluator.py

@@ -6,6 +6,7 @@
 
 from .registry import register_evaluator, BaseEvaluator
 from .whowhat_metrics import TextDivergency, TextSimilarity
+from .utils import patch_awq_for_inference
 
 default_data = {
     "en": {
@@ -187,17 +188,27 @@ def worst_examples(self, top_k: int = 5, metric="similarity"):
 
     def _generate_data(self, model, gen_answer_fn=None, generation_config=None):
         def default_gen_answer(model, tokenizer, prompt, max_new_tokens, crop_question, use_chat_template=False):
+            is_awq = getattr(model, "is_awq", None) is not None
+
             if use_chat_template:
                 message = [{"role": "user", "content": prompt}]
                 inputs = tokenizer.apply_chat_template(message, tokenize=True, add_generation_prompt=True, return_tensors="pt")
-                tokens = model.generate(inputs, do_sample=False, max_new_tokens=max_new_tokens)
+                if is_awq:
+                    with patch_awq_for_inference(is_awq):
+                        tokens = model.generate(inputs, do_sample=False, max_new_tokens=max_new_tokens)
+                else:
+                    tokens = model.generate(inputs, do_sample=False, max_new_tokens=max_new_tokens)
                 if crop_question:
                     tokens = tokens[:, inputs.shape[-1]:]
                 res = self.tokenizer.decode(tokens[0], skip_special_tokens=True)
                 return res
             else:
                 inputs = self.tokenizer(prompt, return_tensors="pt")
-                tokens = model.generate(**inputs, do_sample=False, max_new_tokens=max_new_tokens)
+                if is_awq:
+                    with patch_awq_for_inference(is_awq):
+                        tokens = model.generate(**inputs, do_sample=False, max_new_tokens=max_new_tokens)
+                else:
+                    tokens = model.generate(**inputs, do_sample=False, max_new_tokens=max_new_tokens)
                 if crop_question:
                     tokens = tokens[:, inputs["input_ids"].shape[-1] :]
                 return self.tokenizer.batch_decode(tokens, skip_special_tokens=True)[0]

tools/who_what_benchmark/whowhatbench/utils.py

@@ -1,5 +1,6 @@
 from typing import Union, Tuple, List, Optional
 import torch
+from contextlib import contextmanager
 
 from diffusers.utils import torch_utils
 
@@ -22,3 +23,58 @@ def new_randn_tensor(
 
 def patch_diffusers():
     torch_utils.randn_tensor = new_randn_tensor
+
+
+@contextmanager
+def mock_torch_cuda_is_available(to_patch):
+    original_is_available = torch.cuda.is_available
+    if to_patch:
+        torch.cuda.is_available = lambda: True
+    try:
+        yield
+    finally:
+        if to_patch:
+            torch.cuda.is_available = original_is_available
+
+
+@contextmanager
+def patch_awq_for_inference(to_patch):
+    orig_gemm_forward = None
+    if to_patch:
+        # patch GEMM module to allow inference without CUDA GPU
+        from awq.modules.linear.gemm import WQLinearMMFunction
+        from awq.utils.packing_utils import dequantize_gemm
+
+        def new_forward(
+            ctx,
+            x,
+            qweight,
+            qzeros,
+            scales,
+            w_bit=4,
+            group_size=128,
+            bias=None,
+            out_features=0,
+        ):
+            ctx.out_features = out_features
+
+            out_shape = x.shape[:-1] + (out_features,)
+            x = x.to(torch.float16)
+
+            out = dequantize_gemm(qweight, qzeros, scales, w_bit, group_size)
+            out = torch.matmul(x, out)
+
+            out = out + bias if bias is not None else out
+            out = out.reshape(out_shape)
+
+            if len(out.shape) == 2:
+                out = out.unsqueeze(0)
+            return out
+
+        orig_gemm_forward = WQLinearMMFunction.forward
+        WQLinearMMFunction.forward = new_forward
+    try:
+        yield
+    finally:
+        if orig_gemm_forward is not None:
+            WQLinearMMFunction.forward = orig_gemm_forward