Merge pull request #61 from casper-hansen/bigcode

Add GPT BigCode support (StarCoder)

awq/models/__init__.py

@@ -3,4 +3,5 @@ from .llama import LlamaAWQForCausalLM
 from .opt import OptAWQForCausalLM
 from .falcon import FalconAWQForCausalLM
 from .bloom import BloomAWQForCausalLM
-from .gptj import GPTJAWQForCausalLM
+from .gptj import GPTJAWQForCausalLM
\ No newline at end of file
+from .gpt_bigcode import GptBigCodeAWQForCausalLM
\ No newline at end of file

awq/models/auto.py

@@ -11,7 +11,8 @@ AWQ_CAUSAL_LM_MODEL_MAP = {
     "RefinedWebModel": FalconAWQForCausalLM,
     "falcon": FalconAWQForCausalLM,
     "bloom": BloomAWQForCausalLM,
-    "gptj": GPTJAWQForCausalLM
+    "gptj": GPTJAWQForCausalLM,
+    "gpt_bigcode": GptBigCodeAWQForCausalLM
 }
 
 def check_and_get_model_type(model_dir, trust_remote_code=True):

awq/models/gpt_bigcode.py

+from .base import BaseAWQForCausalLM
+from transformers.models.gpt_bigcode.modeling_gpt_bigcode import GPTBigCodeForCausalLM, GPTBigCodeBlock as OldGptBigCodeBlock
+
+class GptBigCodeAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "GPTBigCodeBlock"
+    max_new_tokens_key = "n_positions"
+
+    @staticmethod
+    def get_model_layers(model: GPTBigCodeForCausalLM):
+        return model.transformer.h
+
+    @staticmethod
+    def get_act_for_scaling(module: OldGptBigCodeBlock):
+        return dict(
+            is_scalable=True,
+            scale_name="mlp.act",
+            scale_layer=module.mlp.act,
+            scale_shape=module.mlp.c_fc.out_features
+        )
+
+    @staticmethod
+    def move_embed(model: GPTBigCodeForCausalLM, device):
+        model.transformer.wte = model.transformer.wte.to(device)
+        model.transformer.drop = model.transformer.drop.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module:OldGptBigCodeBlock, input_feat, module_kwargs):
+        layers = []
+
+        # attention input
+        layers.append(dict(
+            prev_op=module.ln_1,
+            layers=[module.attn.c_attn],
+            inp=input_feat['attn.c_attn'],
+            module2inspect=module.attn,
+            kwargs=module_kwargs
+        ))
+        
+        # linear 1
+        layers.append(dict(
+            prev_op=module.ln_2,
+            layers=[module.mlp.c_fc],
+            inp=input_feat['mlp.c_fc'],
+            module2inspect=module.mlp
+        ))
+
+        # linear 2
+        layers.append(dict(
+            prev_op=module.mlp.act,
+            layers=[module.mlp.c_proj],
+            inp=input_feat['mlp.c_proj']
+        ))
+
+        return layers

awq/quantize/scale.py

 import torch
 import torch.nn as nn
-
 from typing import Tuple
 from awq.modules.act import ScaledActivation
-from transformers.activations import NewGELUActivation
 from awq.utils.module import get_op_by_name, set_op_by_name
 from transformers.models.bloom.modeling_bloom import BloomGelu
 from transformers.models.llama.modeling_llama import LlamaRMSNorm
+from transformers.activations import NewGELUActivation, PytorchGELUTanh
 
 allowed_norms = [nn.LayerNorm, LlamaRMSNorm]
-allowed_act_fns = [nn.GELU, BloomGelu, NewGELUActivation]
+allowed_act_fns = [nn.GELU, BloomGelu, NewGELUActivation, PytorchGELUTanh]
 
 @torch.no_grad()
 def apply_clip(module, clip_list: Tuple[str, torch.Tensor]):