Merge pull request #240 from PanQiWei/support-qwen

support qwen

auto_gptq/modeling/__init__.py

@@ -12,3 +12,4 @@ from .gpt_bigcode import *
 from .codegen import *
 from .baichuan import *
 from .internlm import *
+from .qwen import *

auto_gptq/modeling/_const.py

@@ -20,6 +20,7 @@ SUPPORTED_MODELS = [
     "RefinedWeb",
     "baichuan",
     "internlm",
+    "qwen",
 ]
 if compare_transformers_version("v4.28.0", op="ge"):
     SUPPORTED_MODELS.append("llama")

auto_gptq/modeling/auto.py

@@ -15,6 +15,7 @@ from .rw import RWGPTQForCausalLM
 from .gpt_bigcode import GPTBigCodeGPTQForCausalLM
 from .baichuan import BaiChuanGPTQForCausalLM
 from .internlm import InternLMGPTQForCausalLM
+from .qwen import QwenGPTQForCausalLM
 
 GPTQ_CAUSAL_LM_MODEL_MAP = {
     "bloom": BloomGPTQForCausalLM,
@@ -30,6 +31,7 @@ GPTQ_CAUSAL_LM_MODEL_MAP = {
     "RefinedWeb": RWGPTQForCausalLM,
     "baichuan": BaiChuanGPTQForCausalLM,
     "internlm": InternLMGPTQForCausalLM,
+    "qwen": QwenGPTQForCausalLM,
 }
 
 

auto_gptq/modeling/qwen.py

@@ -0,0 +1,16 @@
+from ._base import *
+
+
+class QwenGPTQForCausalLM(BaseGPTQForCausalLM):
+    layer_type = "QWenBlock"
+    layers_block_name = "transformer.h"
+    outside_layer_modules = ["transformer.wte", "transformer.wpe", "transformer.ln_f"]
+    inside_layer_modules = [
+        ["attn.c_attn"],
+        ["attn.c_proj"],
+        ["mlp.w1", "mlp.w2"],
+        ["mlp.c_proj"]
+    ]
+
+
+__all__ = ["QwenGPTQForCausalLM"]