Merge pull request #240 from PanQiWei/support-qwen

support qwen
AutoGPTQ · Aug 8, 2023 · 3fb7d1e · 3fb7d1e
2 parents 5981f15 + fe24450
commit 3fb7d1e
Show file tree

Hide file tree

Showing 4 changed files with 20 additions and 0 deletions.
diff --git a/auto_gptq/modeling/__init__.py b/auto_gptq/modeling/__init__.py
@@ -12,3 +12,4 @@
 from .codegen import *
 from .baichuan import *
 from .internlm import *
+from .qwen import *
diff --git a/auto_gptq/modeling/_const.py b/auto_gptq/modeling/_const.py
@@ -20,6 +20,7 @@
     "RefinedWeb",
     "baichuan",
     "internlm",
+    "qwen",
 ]
 if compare_transformers_version("v4.28.0", op="ge"):
     SUPPORTED_MODELS.append("llama")

diff --git a/auto_gptq/modeling/auto.py b/auto_gptq/modeling/auto.py
@@ -15,6 +15,7 @@
 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 @@
     "RefinedWeb": RWGPTQForCausalLM,
     "baichuan": BaiChuanGPTQForCausalLM,
     "internlm": InternLMGPTQForCausalLM,
+    "qwen": QwenGPTQForCausalLM,
 }
 
 

diff --git a/auto_gptq/modeling/qwen.py b/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"]