Merge remote-tracking branches 'laaza/Mistral' and 'laaza/MPT'

auto_gptq/modeling/__init__.py

@@ -12,4 +12,6 @@ from .gpt_bigcode import *
 from .codegen import *
 from .baichuan import *
 from .internlm import *
-from .qwen import *
+from .qwen import *
+from .mistral import *
+from .mpt import *

auto_gptq/modeling/_const.py

@@ -21,11 +21,15 @@ SUPPORTED_MODELS = [
     "baichuan",
     "internlm",
     "qwen",
+    "mpt",
 ]
 if compare_transformers_version("v4.28.0", op="ge"):
     SUPPORTED_MODELS.append("llama")
 if compare_transformers_version("v4.33.0", op="ge"):
     SUPPORTED_MODELS.append("falcon")
+if compare_transformers_version("v4.34.0", op="ge"):
+    SUPPORTED_MODELS.append("mistral")
+
 
 EXLLAMA_DEFAULT_MAX_INPUT_LENGTH = 2048
 

auto_gptq/modeling/auto.py

@@ -16,6 +16,8 @@ from .gpt_bigcode import GPTBigCodeGPTQForCausalLM
 from .baichuan import BaiChuanGPTQForCausalLM
 from .internlm import InternLMGPTQForCausalLM
 from .qwen import QwenGPTQForCausalLM
+from .mistral import MistralGPTQForCausalLM
+from .mpt import MPTGPTQForCausalLM
 
 GPTQ_CAUSAL_LM_MODEL_MAP = {
     "bloom": BloomGPTQForCausalLM,
@@ -33,6 +35,8 @@ GPTQ_CAUSAL_LM_MODEL_MAP = {
     "baichuan": BaiChuanGPTQForCausalLM,
     "internlm": InternLMGPTQForCausalLM,
     "qwen": QwenGPTQForCausalLM,
+    "mistral": MistralGPTQForCausalLM,
+    "mpt": MPTGPTQForCausalLM,
 }
 
 

auto_gptq/modeling/mistral.py

@@ -0,0 +1,16 @@
+from ._base import *
+
+
+class MistralGPTQForCausalLM(BaseGPTQForCausalLM):
+    layer_type = "MistralDecoderLayer"
+    layers_block_name = "model.layers"
+    outside_layer_modules = ["model.embed_tokens", "model.norm"]
+    inside_layer_modules = [
+        ["self_attn.k_proj", "self_attn.v_proj", "self_attn.q_proj"],
+        ["self_attn.o_proj"],
+        ["mlp.up_proj", "mlp.gate_proj"],
+        ["mlp.down_proj"],
+    ]
+
+
+__all__ = ["MistralGPTQForCausalLM"]

auto_gptq/modeling/mpt.py

@@ -0,0 +1,18 @@
+from auto_gptq.modeling import BaseGPTQForCausalLM
+
+
+class MPTGPTQForCausalLM(BaseGPTQForCausalLM):
+    layer_type = "MPTBlock"
+    layers_block_name = "transformer.blocks"
+    outside_layer_modules = [
+        "transformer.wte",  "transformer.norm_f"
+    ]
+
+    inside_layer_modules = [
+        ["attn.Wqkv"],
+        ["attn.out_proj"],
+        ["ffn.up_proj"],
+        ["ffn.down_proj"]
+    ]
+
+__all__ = ["MPTGPTQForCausalLM"]