Merge pull request #79 from casper-hansen/mistral

Mistral support

awq/models/__init__.py

@@ -4,4 +4,5 @@ from .opt import OptAWQForCausalLM
 from .falcon import FalconAWQForCausalLM
 from .bloom import BloomAWQForCausalLM
 from .gptj import GPTJAWQForCausalLM
-from .gpt_bigcode import GptBigCodeAWQForCausalLM
+from .gpt_bigcode import GptBigCodeAWQForCausalLM
\ No newline at end of file
+from .mistral import MistralAWQForCausalLM
\ No newline at end of file

awq/models/auto.py

@@ -12,7 +12,8 @@ AWQ_CAUSAL_LM_MODEL_MAP = {
     "falcon": FalconAWQForCausalLM,
     "bloom": BloomAWQForCausalLM,
     "gptj": GPTJAWQForCausalLM,
-    "gpt_bigcode": GptBigCodeAWQForCausalLM
+    "gpt_bigcode": GptBigCodeAWQForCausalLM,
+    "mistral": MistralAWQForCausalLM
 }
 
 def check_and_get_model_type(model_dir, trust_remote_code=True):

awq/models/mistral.py

+import logging
+from .base import BaseAWQForCausalLM
+
+try:
+    from transformers.models.mistral.modeling_mistral import MistralDecoderLayer, MistralForCausalLM
+except:
+    # TODO: Remove once released on PyPi
+    logging.warning("You need the latest transformers 4.34.0.dev0: pip install git+https://github.com/huggingface/transformers.git")
+    MistralForCausalLM = None
+    MistralDecoderLayer = None
+
+class MistralAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "MistralDecoderLayer"
+    max_new_tokens_key = "max_position_embeddings"
+
+    @staticmethod
+    def get_model_layers(model: MistralForCausalLM):
+        return model.model.layers
+    
+    @staticmethod
+    def get_act_for_scaling(module: MistralDecoderLayer):
+        return dict(
+            is_scalable=False
+        )
+    
+    @staticmethod
+    def move_embed(model: MistralForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+    
+    @staticmethod
+    def get_layers_for_scaling(module: MistralDecoderLayer, input_feat, module_kwargs):
+        layers = []
+
+        # attention input
+        layers.append(dict(
+            prev_op=module.input_layernorm,
+            layers=[module.self_attn.q_proj,
+                    module.self_attn.k_proj, module.self_attn.v_proj],
+            inp=input_feat['self_attn.q_proj'],
+            module2inspect=module.self_attn, kwargs=module_kwargs,
+        ))
+
+        # attention out
+        # Please refer to https://github.com/mit-han-lab/llm-awq/pull/67#issue-1850622696
+        if module.self_attn.v_proj.weight.shape == module.self_attn.o_proj.weight.shape:
+            layers.append(dict(
+                prev_op=module.self_attn.v_proj,
+                layers=[module.self_attn.o_proj],
+                inp=input_feat['self_attn.o_proj'],
+            ))
+        
+        # linear 1
+        layers.append(dict(
+            prev_op=module.post_attention_layernorm,
+            layers=[module.mlp.gate_proj, module.mlp.up_proj],
+            inp=input_feat['mlp.gate_proj'],
+            module2inspect=module.mlp,
+        ))
+
+        # linear 2
+        layers.append(dict(
+            prev_op=module.mlp.up_proj,
+            layers=[module.mlp.down_proj],
+            inp=input_feat['mlp.down_proj'],
+        ))
+
+        return layers