Add StableLM support (#410)

Co-authored-by: Casper <casperbh.96@gmail.com>

awq/models/__init__.py

@@ -15,4 +15,5 @@ from .llava import LlavaAWQForCausalLM
 from .mixtral import MixtralAWQForCausalLM
 from .qwen2 import Qwen2AWQForCausalLM
 from .gemma import GemmaAWQForCausalLM
-from .starcoder2 import Starcoder2AWQForCausalLM
\ No newline at end of file
+from .stablelm import StableLmAWQForCausalLM
+from .starcoder2 import Starcoder2AWQForCausalLM

awq/models/auto.py

@@ -24,6 +24,7 @@ AWQ_CAUSAL_LM_MODEL_MAP = {
     "llava": LlavaAWQForCausalLM,
     "qwen2": Qwen2AWQForCausalLM,
     "gemma": GemmaAWQForCausalLM,
+    "stablelm": StableLmAWQForCausalLM,
     "starcoder2": Starcoder2AWQForCausalLM,
 }
 

awq/models/base.py

@@ -68,6 +68,7 @@ TRANSFORMERS_AUTO_MAPPING_DICT = {
     "llava": "AutoModelForVision2Seq",
     "qwen2": "AutoModelForCausalLM",
     "gemma": "AutoModelForCausalLM",
+    "stablelm": "AutoModelForCausalLM",
     "starcoder2": "AutoModelForCausalLM",
 }
 

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/modules/fused/attn.py

@@ -29,9 +29,7 @@ class RoPE(nn.Module):
         super(RoPE, self).__init__()
 
         self.freqs_cis = nn.Parameter(
-            self.precompute_freqs_cis(
-                head_dim, max_seq_len * 2, rope_theta
-            ).to(device),
+            self.precompute_freqs_cis(head_dim, max_seq_len * 2, rope_theta).to(device),
             requires_grad=False,
         )
 
@@ -118,6 +116,7 @@ class QuantAttentionFused(nn.Module):
         use_alibi=False,
         attention_shapes=None,
         rope_theta=10000,
+        partial_rotary_factor=1.0,
         head_dim=None,
         **kwargs
     ):
@@ -127,7 +126,7 @@ class QuantAttentionFused(nn.Module):
         self.n_kv_heads = n_kv_heads
         self.n_kv_groups = n_heads // n_kv_heads if n_kv_heads != 0 else 0
         self.head_dim = head_dim
-        
+
         if head_dim is None:
             self.head_dim = hidden_size // n_heads
 
@@ -167,8 +166,9 @@ class QuantAttentionFused(nn.Module):
             self.is_neox = False
         else:
             self.alibi = None
-            self.rope = RoPE(self.head_dim, max_seq_len, dev, rope_theta)
-            self.rotary_dim = self.head_dim
+            self.partial_rotary_factor = partial_rotary_factor
+            self.rotary_dim = int(self.head_dim * self.partial_rotary_factor)
+            self.rope = RoPE(self.rotary_dim, max_seq_len, dev, rope_theta)
             self.is_neox = True
 
     def forward(
@@ -209,13 +209,27 @@ class QuantAttentionFused(nn.Module):
         xk = self.attention_shapes["xk_slice"](xqkv)
         xv = self.attention_shapes["xv_slice"](xqkv)
 
-        if seqlen > 1 or not FT_INSTALLED:
+        if seqlen > 1 or self.partial_rotary_factor < 1 or not FT_INSTALLED:
             xq = xq.view((bsz, seqlen) + self.attention_shapes["xq_view"])
             xk = xk.view((bsz, seqlen) + self.attention_shapes["xk_view"])
             xv = xv.view((bsz, seqlen) + self.attention_shapes["xv_view"])
 
             if not self.use_alibi:
-                xq, xk = self.rope.forward(xq, xk, self.start_pos, seqlen)
+                # Partial rotary embedding
+                if self.partial_rotary_factor < 1:
+                    xq_rot, xq_pass = (
+                        xq[..., : self.rotary_dim],
+                        xq[..., self.rotary_dim :],
+                    )
+                    xk_rot, xk_pass = (
+                        xk[..., : self.rotary_dim],
+                        xk[..., self.rotary_dim :],
+                    )
+                    xq_rot, xk_rot = self.rope.forward(xq_rot, xk_rot, self.start_pos, seqlen)
+                    xq = torch.cat((xq_rot, xq_pass), dim=-1)
+                    xk = torch.cat((xk_rot, xk_pass), dim=-1)
+                else:
+                    xq, xk = self.rope.forward(xq, xk, self.start_pos, seqlen)
 
             values_store = xv.transpose(2, 1)
             keys_store = (

awq/modules/fused/block.py

@@ -79,6 +79,7 @@ class LlamaLikeBlock(nn.Module):
         dev,
         max_seq_len,
         rope_theta=10000,
+        partial_rotary_factor=1.0,
         use_alibi=False,
         head_dim=None,
     ):
@@ -103,6 +104,7 @@ class LlamaLikeBlock(nn.Module):
             max_seq_len=max_seq_len,
             use_alibi=use_alibi,
             rope_theta=rope_theta,
+            partial_rotary_factor=partial_rotary_factor,
             head_dim=head_dim,
         ).to(dev)
         self.norm_2 = norm_2.to(dev)