Merge branch 'v0.2.5_develop' into 'main'

v0.2.5

See merge request dcutoolkit/deeplearing/autoawq!2

awq/models/gemma.py

+import tqdm
+import torch
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.gemma.modeling_gemma import (
+    GemmaDecoderLayer as OldGemmaDecoderLayer,
+    GemmaForCausalLM as OldGemmaForCausalLM,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class GemmaAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "GemmaDecoderLayer"
+    max_new_tokens_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldGemmaDecoderLayer):
+        fuser = GemmaFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldGemmaForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldGemmaDecoderLayer):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldGemmaForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: OldGemmaDecoderLayer, 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
+
+
+class GemmaFuser:
+    def __init__(self, model: OldGemmaForCausalLM):
+        self.model = model
+
+        self.Gemma_blocks: List[Tuple[str, OldGemmaDecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "GemmaDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldGemmaDecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            with torch.no_grad():
+                # GemmaRMSNorm is different from Llama's in that it multiplies
+                # (1 + weight) to the output, instead of just weight.
+                module.input_layernorm.weight += 1
+                module.post_attention_layernorm.weight += 1
+            norm_1 = FasterTransformerRMSNorm(
+                module.input_layernorm.weight, module.input_layernorm.eps
+            )
+            norm_2 = FasterTransformerRMSNorm(
+                module.post_attention_layernorm.weight,
+                module.post_attention_layernorm.eps,
+            )
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                    rope_theta=self.model.config.rope_theta,
+                    head_dim=self.model.config.head_dim,
+                )
+            )
+        
+        with torch.no_grad():
+            # Normalize Gemma's embedding layer
+            self.model.model.embed_tokens.weight *= self.model.config.hidden_size**0.5
+        
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

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_seq_len_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.wpe = model.transformer.wpe.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/models/gpt_neox.py

+from .base import BaseAWQForCausalLM
+from transformers.models.gpt_neox.modeling_gpt_neox import (
+    GPTNeoXLayer,
+    GPTNeoXForCausalLM,
+)
+
+
+class GPTNeoXAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "GPTNeoXDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def get_model_layers(model: GPTNeoXForCausalLM):
+        return model.gpt_neox.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: GPTNeoXLayer):
+        return dict(
+            is_scalable=True,
+            scale_name="mlp.act",
+            scale_layer=module.mlp.act,
+            scale_shape=module.mlp.dense_h_to_4h.out_features,
+        )
+
+    @staticmethod
+    def move_embed(model: GPTNeoXForCausalLM, device: str):
+        model.gpt_neox.embed_in = model.gpt_neox.embed_in.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: GPTNeoXLayer, input_feat, module_kwargs):
+        layers = []
+
+        # attention input
+        layers.append(
+            dict(
+                prev_op=module.input_layernorm,
+                layers=[module.attention.query_key_value],
+                inp=input_feat["attention.query_key_value"],
+            )
+        )
+
+        # attention out
+        # Please refer to https://github.com/mit-han-lab/llm-awq/issues/2#issuecomment-1606297469
+        """
+        layers.append(dict(
+            prev_op=module.attention.query_key_value,
+            layers=[module.attention.dense],
+            inp=input_feat['attention.dense'],
+        ))
+        """
+
+        # linear 1
+        layers.append(
+            dict(
+                prev_op=module.post_attention_layernorm,
+                layers=[module.mlp.dense_h_to_4h],
+                inp=input_feat["mlp.dense_h_to_4h"],
+            )
+        )
+
+        # linear 2
+        layers.append(
+            dict(
+                prev_op=module.mlp.act,
+                layers=[module.mlp.dense_4h_to_h],
+                inp=input_feat["mlp.dense_4h_to_h"],
+            )
+        )
+
+        return layers

awq/models/gptj.py

+from .base import BaseAWQForCausalLM
+from transformers.models.gptj.modeling_gptj import GPTJForCausalLM, GPTJBlock
+
+
+class GPTJAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "GPTJBlock"
+    max_seq_len_key = "n_positions"
+
+    @staticmethod
+    def get_model_layers(model: GPTJForCausalLM):
+        return model.transformer.h
+
+    @staticmethod
+    def get_act_for_scaling(module: GPTJBlock):
+        return dict(
+            is_scalable=True,
+            scale_name="mlp.act",
+            scale_layer=module.mlp.act,
+            scale_shape=module.mlp.fc_in.out_features,
+        )
+
+    @staticmethod
+    def move_embed(model: GPTJForCausalLM, device: str):
+        model.transformer.wte = model.transformer.wte.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: GPTJBlock, input_feat, module_kwargs):
+        layers = []
+
+        # attention input + linear 1
+        layers.append(
+            dict(
+                prev_op=module.ln_1,
+                layers=[
+                    module.attn.q_proj,
+                    module.attn.k_proj,
+                    module.attn.v_proj,
+                    module.mlp.fc_in,
+                ],
+                inp=input_feat["attn.q_proj"],
+                module2inspect=module,
+                kwargs=module_kwargs,
+            )
+        )
+
+        # attention out
+        layers.append(
+            dict(
+                prev_op=module.attn.v_proj,
+                layers=[module.attn.out_proj],
+                inp=input_feat["attn.out_proj"],
+            )
+        )
+
+        # linear 2
+        layers.append(
+            dict(
+                prev_op=module.mlp.act,
+                layers=[module.mlp.fc_out],
+                inp=input_feat["mlp.fc_out"],
+            )
+        )
+
+        return layers

awq/models/llama.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.llama.modeling_llama import (
+    LlamaDecoderLayer as OldLlamaDecoderLayer,
+    LlamaForCausalLM as OldLlamaForCausalLM,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class LlamaAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "LlamaDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldLlamaForCausalLM):
+        fuser = LlamaFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldLlamaForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldLlamaDecoderLayer):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldLlamaForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: OldLlamaDecoderLayer, 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
+
+
+class LlamaFuser:
+    def __init__(self, model: OldLlamaForCausalLM):
+        self.model = model
+
+        self.llama_blocks: List[Tuple[str, OldLlamaDecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "LlamaDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldLlamaDecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = FasterTransformerRMSNorm(
+                module.input_layernorm.weight, module.input_layernorm.variance_epsilon
+            )
+            norm_2 = FasterTransformerRMSNorm(
+                module.post_attention_layernorm.weight,
+                module.post_attention_layernorm.variance_epsilon,
+            )
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                    rope_theta=self.model.config.rope_theta,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/models/llava.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.llama.modeling_llama import (
+    LlamaDecoderLayer as OldLlamaDecoderLayer,
+)
+from transformers.models.llava.modeling_llava import (
+    LlavaForConditionalGeneration as OldLlavaForConditionalGeneration,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class LlavaAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "LlamaDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldLlavaForConditionalGeneration):
+        fuser = LlavaFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldLlavaForConditionalGeneration):
+        return model.language_model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldLlamaDecoderLayer):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldLlavaForConditionalGeneration, device: str):
+        model.language_model.model.embed_tokens = model.get_input_embeddings().to(
+            device
+        )
+
+    @staticmethod
+    def get_layers_for_scaling(module: OldLlamaDecoderLayer, 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
+
+
+class LlavaFuser:
+    def __init__(self, model: OldLlavaForConditionalGeneration):
+        self.model = model.language_model
+
+        self.llama_blocks: List[Tuple[str, OldLlamaDecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "LlamaDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldLlamaDecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = FasterTransformerRMSNorm(
+                module.input_layernorm.weight, module.input_layernorm.variance_epsilon
+            )
+            norm_2 = FasterTransformerRMSNorm(
+                module.post_attention_layernorm.weight,
+                module.post_attention_layernorm.variance_epsilon,
+            )
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/models/mistral.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.mistral.modeling_mistral import (
+    MistralDecoderLayer as OldMistralDecoderLayer,
+    MistralForCausalLM as OldMistralForCausalLM,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class MistralAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "MistralDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldMistralForCausalLM):
+        fuser = MistralFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldMistralForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldMistralDecoderLayer):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldMistralForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(
+        module: OldMistralDecoderLayer, 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
+
+
+class MistralFuser:
+    def __init__(self, model: OldMistralForCausalLM):
+        self.model = model
+
+        self.mistral_blocks: List[Tuple[str, OldMistralDecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "MistralDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldMistralDecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = FasterTransformerRMSNorm(
+                module.input_layernorm.weight, module.input_layernorm.variance_epsilon
+            )
+            norm_2 = FasterTransformerRMSNorm(
+                module.post_attention_layernorm.weight,
+                module.post_attention_layernorm.variance_epsilon,
+            )
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/models/mixtral.py

+import tqdm
+import torch
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.modules.fused.block import MixtralBlock
+from awq.modules.fused.model import MixtralModel
+from awq.modules.fused.moe import FusedSparseMoeBlock
+from awq.utils.fused_utils import fuse_qkv, fuse_linears
+from transformers.models.mixtral.modeling_mixtral import (
+    MixtralDecoderLayer as OldMixtralDecoderLayer,
+    MixtralForCausalLM as OldMixtralForCausalLM,
+)
+from awq.modules.linear import WQLinear_GEMM
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class MixtralAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "MixtralDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+    modules_to_not_convert = ["gate"]
+
+    @staticmethod
+    def fuse_layers(model: OldMixtralForCausalLM):
+        fuser = MixtralFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldMixtralForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldMixtralForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(
+        module: OldMixtralDecoderLayer, 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
+        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 in
+        layers.append(
+            dict(
+                prev_op=module.post_attention_layernorm,
+                layers=[
+                    w
+                    for expert in module.block_sparse_moe.experts
+                    for w in [expert.w1, expert.w3]
+                ],
+                inp=input_feat["block_sparse_moe"],
+                module2inspect=module.block_sparse_moe,
+            )
+        )
+
+        # linear out
+        for i, expert in enumerate(module.block_sparse_moe.experts):
+            layers.append(
+                dict(
+                    prev_op=expert.w3,
+                    layers=[expert.w2],
+                    inp=input_feat[f"block_sparse_moe.experts.{i}.w2"],
+                )
+            )
+
+        return layers
+
+
+class MixtralFuser:
+    def __init__(self, model: OldMixtralForCausalLM):
+        self.model = model
+
+        self.mixtral_blocks: List[Tuple[str, OldMixtralDecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "MixtralDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldMixtralDecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = FasterTransformerRMSNorm(
+                module.input_layernorm.weight, module.input_layernorm.variance_epsilon
+            )
+
+            norm_2 = FasterTransformerRMSNorm(
+                module.post_attention_layernorm.weight,
+                module.post_attention_layernorm.variance_epsilon,
+            )
+
+            sparse_moe = module.block_sparse_moe
+            if isinstance(sparse_moe.experts[0].w1, WQLinear_GEMM):
+                fused_w1w3s = [
+                    fuse_linears(
+                        [
+                            sparse_moe.experts[i].w1,
+                            sparse_moe.experts[i].w3,
+                        ],
+                        device,
+                    )
+                    for i in range(len(sparse_moe.experts))
+                ]
+
+                stacked_w1w3s = fuse_linears(
+                    fused_w1w3s, device, dim=0, operation=torch.stack
+                )
+
+                stacked_w2s = fuse_linears(
+                    [expert.w2 for expert in sparse_moe.experts],
+                    device,
+                    dim=0,
+                    operation=torch.stack,
+                )
+
+                sparse_moe = FusedSparseMoeBlock(
+                    top_k=sparse_moe.top_k,
+                    gate=sparse_moe.gate,
+                    ws=stacked_w1w3s,
+                    w2s=stacked_w2s,
+                )
+
+            blocks.append(
+                MixtralBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    moe=sparse_moe,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                    rope_theta=self.model.config.rope_theta,
+                )
+            )
+        
+        model_norm = FasterTransformerRMSNorm(
+            self.model.model.norm.weight,
+            self.model.model.norm.variance_epsilon,
+        )
+
+        self.model.model = MixtralModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            model_norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/models/mpt.py

+from .base import BaseAWQForCausalLM
+from transformers.models.mpt.modeling_mpt import MptBlock as OldMptBlock, MptForCausalLM
+
+
+class MptAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "MPTBlock"
+    max_seq_len_key = "max_seq_len"
+
+    @staticmethod
+    def fuse_layers(model: MptForCausalLM):
+        fuser = MptFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: MptForCausalLM):
+        return model.transformer.blocks
+
+    @staticmethod
+    def get_act_for_scaling(module: OldMptBlock):
+        return dict(
+            is_scalable=True,
+            scale_name="ffn.act",
+            scale_layer=module.ffn.act,
+            scale_shape=module.ffn.up_proj.out_features,
+        )
+
+    @staticmethod
+    def move_embed(model: MptForCausalLM, device: str):
+        model.transformer.wte = model.transformer.wte.to(device)
+        model.transformer.emb_drop = model.transformer.emb_drop.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: OldMptBlock, input_feat, module_kwargs):
+        layers = []
+
+        if module_kwargs.get("output_attentions") is not None:
+            module_kwargs.pop("output_attentions")
+
+        # attention input
+        layers.append(
+            dict(
+                prev_op=module.norm_1,
+                layers=[module.attn.Wqkv],
+                inp=input_feat["attn.Wqkv"],
+                module2inspect=module.attn,
+                kwargs=module_kwargs,
+            )
+        )
+
+        # attention output
+        layers.append(
+            dict(
+                prev_op=module.attn.Wqkv,
+                layers=[module.attn.out_proj],
+                inp=input_feat["attn.out_proj"],
+            )
+        )
+
+        # linear 1
+        layers.append(
+            dict(
+                prev_op=module.norm_2,
+                layers=[module.ffn.up_proj],
+                inp=input_feat["ffn.up_proj"],
+                module2inspect=module.ffn,
+            )
+        )
+
+        # linear 2
+        layers.append(
+            dict(
+                prev_op=module.ffn.act,
+                layers=[module.ffn.down_proj],
+                inp=input_feat["ffn.down_proj"],
+            )
+        )
+
+        return layers
+
+
+from typing import List, Tuple
+from awq.utils.utils import set_module_name
+from awq.modules.fused.block import MPTBlock
+from awq.modules.fused.model import MPTModel
+
+
+class MptFuser:
+    def __init__(self, model: MptForCausalLM):
+        self.model = model
+
+        self.mpt_blocks: List[Tuple[str, OldMptBlock]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "mptblock" in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldMptBlock
+        for module in self.model.transformer.blocks:
+            blocks.append(
+                MPTBlock(
+                    self.model.config.d_model,
+                    self.model.config.n_heads,
+                    module.attn.Wqkv,
+                    module.attn.out_proj,
+                    module.ffn,
+                    module.norm_1,
+                    module.norm_2,
+                    next(iter(module.state_dict().values())).device,
+                    self.model.config.max_seq_len,
+                )
+            )
+
+        self.model.transformer = MPTModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.transformer.wte,
+            self.model.transformer.norm_f,
+        )
+
+        setattr(self.model.transformer, "blocks", self.model.transformer.blocks)

awq/models/opt.py

+from .base import BaseAWQForCausalLM
+from transformers.models.opt.modeling_opt import OPTForCausalLM, OPTDecoderLayer
+
+
+class OptAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "OPTDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def get_model_layers(model: OPTForCausalLM):
+        return model.model.decoder.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OPTDecoderLayer):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OPTForCausalLM, device: str):
+        model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.to(device)
+        model.model.decoder.embed_positions = model.model.decoder.embed_positions.to(
+            device
+        )
+
+    @staticmethod
+    def get_layers_for_scaling(module: OPTDecoderLayer, input_feat, module_kwargs):
+        layers = []
+
+        # attention input
+        layers.append(
+            dict(
+                prev_op=module.self_attn_layer_norm,
+                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
+        layers.append(
+            dict(
+                prev_op=module.self_attn.v_proj,
+                layers=[module.self_attn.out_proj],
+                inp=input_feat["self_attn.out_proj"],
+            )
+        )
+
+        # linear 1
+        layers.append(
+            dict(
+                prev_op=module.final_layer_norm,
+                layers=[module.fc1],
+                inp=input_feat["fc1"],
+            )
+        )
+
+        # linear 2
+        layers.append(
+            dict(
+                prev_op=module.fc1,
+                layers=[module.fc2],
+                inp=input_feat["fc2"],
+            )
+        )
+
+        return layers

awq/models/qwen.py

+from .base import BaseAWQForCausalLM
+
+
+class QwenAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "QWenBlock"
+    max_seq_len_key = "seq_length"
+
+    @staticmethod
+    def get_model_layers(model):
+        return model.transformer.h
+
+    @staticmethod
+    def get_act_for_scaling(module):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model, device: str):
+        model.transformer.wte = model.transformer.wte.to(device)
+        model.transformer.rotary_emb = model.transformer.rotary_emb.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module, input_feat, module_kwargs):
+        layers = []
+
+        # attention
+        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,
+            )
+        )
+
+        # mlp
+        layers.append(
+            dict(
+                prev_op=module.ln_2,
+                layers=[module.mlp.w2, module.mlp.w1],
+                inp=input_feat["mlp.w2"],
+                module2inspect=module.mlp,
+            )
+        )
+
+        # linear 2
+        layers.append(
+            dict(
+                prev_op=module.mlp.w1,
+                layers=[module.mlp.c_proj],
+                inp=input_feat["mlp.c_proj"],
+            )
+        )
+
+        return layers

awq/models/qwen2.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.qwen2.modeling_qwen2 import (
+    Qwen2DecoderLayer as OldQwen2DecoderLayer,
+    Qwen2ForCausalLM as OldQwen2ForCausalLM,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class Qwen2AWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "Qwen2DecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldQwen2ForCausalLM):
+        fuser = Qwen2Fuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldQwen2ForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldQwen2DecoderLayer):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldQwen2ForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: OldQwen2DecoderLayer, 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
+
+
+class Qwen2Fuser:
+    def __init__(self, model: OldQwen2ForCausalLM):
+        self.model = model
+
+        self.qwen2_blocks: List[Tuple[str, OldQwen2DecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "Qwen2DecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldQwen2DecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = FasterTransformerRMSNorm(
+                module.input_layernorm.weight, module.input_layernorm.variance_epsilon
+            )
+            norm_2 = FasterTransformerRMSNorm(
+                module.post_attention_layernorm.weight,
+                module.post_attention_layernorm.variance_epsilon,
+            )
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/models/stablelm.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.stablelm import StableLmForCausalLM as OldStableLmForCausalLM
+from transformers.models.stablelm.modeling_stablelm import (
+    StableLmDecoderLayer as OldStableLmDecoderLayer,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class StableLmAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "StableLmDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldStableLmForCausalLM):
+        fuser = StableLmFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldStableLmForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldStableLmForCausalLM):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldStableLmForCausalLM, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(
+        module: OldStableLmDecoderLayer, 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
+
+
+class StableLmFuser:
+    def __init__(self, model: OldStableLmForCausalLM):
+        self.model = model
+
+        self.stablelm_blocks: List[Tuple[str, OldStableLmDecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "StableLmDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldStableLmDecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = module.input_layernorm
+            norm_2 = module.post_attention_layernorm
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                    rope_theta=self.model.config.rope_theta,
+                    partial_rotary_factor=self.model.config.partial_rotary_factor,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/models/starcoder2.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from transformers.models.starcoder2.modeling_starcoder2 import (
+    Starcoder2ForCausalLM as OldStarcoder2ForCausalLM,
+    Starcoder2DecoderLayer as OldStarcoder2DecoderLayer,
+)
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class Starcoder2AWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "Starcoder2DecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model: OldStarcoder2ForCausalLM):
+        fuser = Starcoder2Fuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model: OldStarcoder2ForCausalLM):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module: OldStarcoder2DecoderLayer):
+        return dict(
+            is_scalable=True,
+            scale_name="mlp.act",
+            scale_layer=module.mlp.act,
+            scale_shape=module.mlp.c_fc.out_features,
+        )
+        # return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model: OldStarcoder2ForCausalLM, device):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module: OldStarcoder2DecoderLayer, 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
+        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.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
+
+class Starcoder2Fuser:
+    def __init__(self, model: OldStarcoder2ForCausalLM):
+        self.model = model
+
+        self.starcoder2_blocks: List[Tuple[str, OldStarcoder2DecoderLayer]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "Starcoder2DecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        module: OldStarcoder2DecoderLayer
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            # SC2 use normal LayerNorm
+            norm_1 = module.input_layernorm
+            norm_2 = module.post_attention_layernorm
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)
\ No newline at end of file

awq/models/yi.py

+import tqdm
+from typing import List, Tuple
+from .base import BaseAWQForCausalLM
+from awq.utils.fused_utils import fuse_qkv
+from awq.modules.fused.block import LlamaLikeBlock
+from awq.modules.fused.model import LlamaLikeModel
+from awq.modules.fused.norm import FasterTransformerRMSNorm
+
+
+class YiAWQForCausalLM(BaseAWQForCausalLM):
+    layer_type = "YiDecoderLayer"
+    max_seq_len_key = "max_position_embeddings"
+
+    @staticmethod
+    def fuse_layers(model):
+        fuser = YiFuser(model)
+        fuser.fuse_transformer()
+
+    @staticmethod
+    def get_model_layers(model):
+        return model.model.layers
+
+    @staticmethod
+    def get_act_for_scaling(module):
+        return dict(is_scalable=False)
+
+    @staticmethod
+    def move_embed(model, device: str):
+        model.model.embed_tokens = model.model.embed_tokens.to(device)
+
+    @staticmethod
+    def get_layers_for_scaling(module, input_feat, module_kwargs):
+        layers = []
+
+        # attention input
+        layers.append(
+            dict(
+                prev_op=module.ln1,
+                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.ln2,
+                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
+
+
+class YiFuser:
+    def __init__(self, model):
+        self.model = model
+
+        self.yi_blocks: List[Tuple[str, object]] = [
+            (name, module)
+            for name, module in self.model.named_modules()
+            if "YiDecoderLayer".lower() in module.__class__.__name__.lower()
+        ]
+
+    def fuse_transformer(self):
+        blocks = []
+
+        for module in tqdm.tqdm(self.model.model.layers, desc="Fusing layers..."):
+            device = next(iter(module.state_dict().values())).device
+            qkv = fuse_qkv(
+                module,
+                module.self_attn.q_proj,
+                module.self_attn.k_proj,
+                module.self_attn.v_proj,
+            )
+            norm_1 = FasterTransformerRMSNorm(
+                module.ln1.weight, module.ln1.variance_epsilon
+            )
+            norm_2 = FasterTransformerRMSNorm(
+                module.ln2.weight, module.ln2.variance_epsilon
+            )
+            blocks.append(
+                LlamaLikeBlock(
+                    hidden_size=self.model.config.hidden_size,
+                    n_heads=self.model.config.num_attention_heads,
+                    n_kv_heads=self.model.config.num_key_value_heads,
+                    qkv_layer=qkv,
+                    o_proj=module.self_attn.o_proj,
+                    mlp=module.mlp,
+                    norm_1=norm_1,
+                    norm_2=norm_2,
+                    dev=device,
+                    max_seq_len=self.model.config.max_seq_len,
+                    rope_theta=self.model.config.rope_theta,
+                )
+            )
+
+        self.model.model = LlamaLikeModel(
+            self.model.config.vocab_size,
+            blocks,
+            self.model.model.embed_tokens,
+            self.model.model.norm,
+        )
+        setattr(self.model.model, "blocks", self.model.model.blocks)

awq/modules/act.py

+import torch.nn as nn
+
+
+class ScaledActivation(nn.Module):
+    def __init__(self, module, scales):
+        super().__init__()
+        self.act = module
+        self.scales = nn.Parameter(scales.data)
+
+    def forward(self, x):
+        return self.act(x) / self.scales.view(1, 1, -1).to(x.device)