Merge branch 'v0.15.1-dev-wm-1' into 'v0.15.1-dev'

[perf]glm4_moe模型适配rmsquant和silu_quant融合算子

See merge request dcutoolkit/deeplearing/vllm!467

vllm/envs.py

@@ -281,6 +281,7 @@ if TYPE_CHECKING:
     VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
     VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
     USE_FUSED_RMS_QUANT: bool = False
+    USE_FUSED_SILU_MUL_QUANT: bool = False
     VLLM_USE_PD_SPLIT: bool = False
     VLLM_USE_PP_SYNC: bool = False
     VLLM_USE_PIECEWISE: bool = False
@@ -1804,6 +1805,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # vllm will use rmsquant fused op
     "USE_FUSED_RMS_QUANT":
         lambda: bool(int(os.getenv("USE_FUSED_RMS_QUANT", "0"))),
+        
+    # vllm will use silu_mul_quant fused op
+    "USE_FUSED_SILU_MUL_QUANT":
+    lambda: bool(int(os.getenv("USE_FUSED_SILU_MUL_QUANT", "0"))),
+
     # vLLM will split prefill and decode, not mix up
     "VLLM_USE_PD_SPLIT":
         lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "True").lower() in

vllm/model_executor/layers/layernorm.py

@@ -6,6 +6,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+from typing import Optional
 from vllm._aiter_ops import rocm_aiter_ops
 from vllm.model_executor.custom_op import CustomOp
 
@@ -14,6 +15,7 @@ from vllm.model_executor.layers.batch_invariant import (
     vllm_is_batch_invariant,
 )
 from vllm.platforms import current_platform
+from vllm.utils import direct_register_custom_op
 from vllm import envs
 
 
@@ -298,6 +300,96 @@ class RMSNorm(CustomOp):
         return s
     
 
+class FusedRMSNormQuant(nn.Module):
+    """Fuse Root mean square normalization and int8 quant.
+
+    Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
+    Refer to https://arxiv.org/abs/1910.07467
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-6,
+        var_hidden_size: int | None = None,
+        has_weight: bool = True,
+        dtype: torch.dtype | None = None,
+    ) -> None:
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.variance_epsilon = eps
+        self.variance_size_override = (
+            None if var_hidden_size == hidden_size else var_hidden_size
+        )
+        weight_dtype = dtype or torch.get_default_dtype()
+        self.has_weight = has_weight
+        self.weight = torch.ones(hidden_size, dtype=weight_dtype)
+        if self.has_weight:
+            self.weight = nn.Parameter(self.weight)
+    
+    def forward(
+        self,
+        x: torch.Tensor,
+        residual: torch.Tensor | None = None,
+        quant_dtype: torch.dtype = torch.int8,
+        update_input: Optional[bool] = True
+    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+        i_q, i_s = torch.ops.vllm.fused_rmsquant(input=x,
+                                                 weight=self.weight,
+                                                 epsilon=self.variance_epsilon,
+                                                 quant_dtype=quant_dtype,
+                                                 residual=residual,
+                                                 update_input=update_input)
+
+        return i_q, i_s, residual
+    
+
+def fused_rmsquant_impl(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    epsilon: float,
+    quant_dtype: torch.dtype,
+    residual: Optional[torch.Tensor] = None,
+    update_input: Optional[bool] = True
+) -> tuple[torch.Tensor, torch.Tensor]:
+    output = torch.empty_like(input, device=input.device, dtype=quant_dtype)
+    scales = torch.empty((input.numel() // input.shape[-1], 1),
+                        device=input.device,
+                        dtype=torch.float32)
+    
+    from lightop.op import rms_norm_dynamic_per_token_quant as ligtop_rms_norm_dynamic_per_token_quant
+    ligtop_rms_norm_dynamic_per_token_quant(output, input, weight,
+                                               scales, epsilon,
+                                               residual, update_input)
+    return output, scales
+
+def fused_rmsquant_fake(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    epsilon: float,
+    quant_dtype: torch.dtype,
+    residual: Optional[torch.Tensor] = None,
+    update_input: Optional[bool] = True
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Fake implementation for torch.compile"""
+    output = torch.empty_like(input, dtype=quant_dtype)
+    scales = torch.empty((input.numel() // input.shape[-1], 1),
+                        device=input.device,
+                        dtype=torch.float32)
+    return output, scales
+
+# from torch.library import Library
+# customer_lib = Library("customer_", "FRAGMENT")
+
+direct_register_custom_op(
+    op_name="fused_rmsquant",
+    op_func=fused_rmsquant_impl,
+    mutates_args=[],
+    fake_impl=fused_rmsquant_fake,
+)
+
+
 # --8<-- [start:gemma_rms_norm]
 @CustomOp.register("gemma_rms_norm")
 class GemmaRMSNorm(CustomOp):

vllm/model_executor/layers/linear.py

@@ -654,11 +654,16 @@ class ColumnParallelLinear(LinearBase):
     def forward(
         self,
         input_,
+        *,
+        iqis: tuple[torch.Tensor, torch.Tensor] | None = None
     ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
         bias = self.bias if not self.skip_bias_add else None
 
         # Matrix multiply.
         assert self.quant_method is not None
+        if iqis is not None:
+            output_parallel = self.quant_method.apply(self, input_, bias, input_quant_args=iqis)
+        else:
             output_parallel = self.quant_method.apply(self, input_, bias)
 
         if self.gather_output and self.tp_size > 1:
@@ -1523,6 +1528,8 @@ class RowParallelLinear(LinearBase):
     def forward(
         self,
         input_,
+        *,
+        iqis: tuple[torch.Tensor, torch.Tensor] | None = None
     ) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
         if self.input_is_parallel:
             input_parallel = input_
@@ -1537,6 +1544,9 @@ class RowParallelLinear(LinearBase):
         # Only fuse bias add into GEMM for rank 0 (this ensures that
         # bias will not get added more than once in TP>1 case)
         bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
+        if iqis is not None:
+            output_parallel = self.quant_method.apply(self, input_parallel, bias_, input_quant_args=iqis)
+        else:
             output_parallel = self.quant_method.apply(self, input_parallel, bias_)
 
         if self.reduce_results and self.tp_size > 1:

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -7,6 +7,7 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.platforms import current_platform
 from vllm.utils import W8a8GetCacheJSON
+from vllm import envs
 
 try:
     from lmslim.layers.gemm.int8_utils import per_token_quant_int8
@@ -167,6 +168,15 @@ def apply_int8_linear(
     # ops.scaled_int8_quant supports both dynamic and static quant.
     # * dynamic, layer.input_scale is None and x_scale computed from x.
     # * static, layer.input_scale is scalar and x_scale is input_scale.
+    if envs.USE_FUSED_RMS_QUANT and input_quant_args is not None:
+        assert len(input_quant_args) == 2
+        x_zp =None
+        x_q, x_scale = input_quant_args
+    elif envs.USE_FUSED_RMS_QUANT and silu_quant_args is not None:
+        assert len(silu_quant_args) == 2
+        x_zp =None
+        x_q, x_scale = silu_quant_args
+    else:
         symmetric = azp_adj is None
         if input_scale is None and input_zero_point is None and symmetric is True:
             x_q, x_scale=per_token_quant_int8(input)

vllm/model_executor/models/glm4_moe.py

@@ -44,7 +44,7 @@ from vllm.distributed import (
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.fused_moe import SharedFusedMoE
-from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.layernorm import RMSNorm, FusedRMSNormQuant
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
     QKVParallelLinear,
@@ -108,8 +108,14 @@ class Glm4MoeMLP(nn.Module):
             )
         self.act_fn = SiluAndMul()
 
-    def forward(self, x):
-        gate_up, _ = self.gate_up_proj(x)
+    def forward(self, x, 
+                iqis: tuple[torch.Tensor, torch.Tensor] | None = None):
+        gate_up, _ = self.gate_up_proj(x, iqis=iqis)
+        if envs.USE_FUSED_SILU_MUL_QUANT:
+            from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
+            xq, xs = lm_fuse_silu_mul_quant(gate_up)
+            x, _ = self.down_proj(gate_up, iqis=(xq, xs))
+        else:
             x = self.act_fn(gate_up)
             x, _ = self.down_proj(x)
         return x
@@ -321,8 +327,9 @@ class Glm4MoeAttention(nn.Module):
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
+        iqis: tuple[torch.Tensor, torch.Tensor] | None = None
     ) -> torch.Tensor:
-        qkv, _ = self.qkv_proj(hidden_states)
+        qkv, _ = self.qkv_proj(hidden_states, iqis=iqis)
 
         if not envs.VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE:
             q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
@@ -408,7 +415,16 @@ class Glm4MoeDecoderLayer(nn.Module):
                 prefix=f"{prefix}.mlp",
             )
 
+        if envs.USE_FUSED_RMS_QUANT:
+            self.input_layernorm = FusedRMSNormQuant(config.hidden_size, eps=config.rms_norm_eps)
+        else:
             self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        if envs.USE_FUSED_RMS_QUANT and isinstance(self.mlp, Glm4MoeMLP):
+            self.post_attention_layernorm = FusedRMSNormQuant(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+        else:
             self.post_attention_layernorm = RMSNorm(
                 config.hidden_size, eps=config.rms_norm_eps
             )
@@ -420,12 +436,33 @@ class Glm4MoeDecoderLayer(nn.Module):
         hidden_states: torch.Tensor,
         residual: torch.Tensor | None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
+        if not envs.USE_FUSED_RMS_QUANT:
             if residual is None:
                 residual = hidden_states
                 hidden_states = self.input_layernorm(hidden_states)
             else:
                 hidden_states, residual = self.input_layernorm(hidden_states, residual)
             hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
+        else:
+            if residual is None:
+                residual = hidden_states.clone()
+                i_q, i_s, _ = self.input_layernorm(x=hidden_states, 
+                                                residual=None, 
+                                                quant_dtype=torch.int8,
+                                                update_input=False
+                                                )
+            else:
+                i_q, i_s, residual = self.input_layernorm(x=hidden_states, 
+                                                      residual=residual, 
+                                                      quant_dtype=torch.int8,
+                                                      update_input=False
+                                                      )
+            hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states, iqis=(i_q, i_s))
+        
+        if envs.USE_FUSED_RMS_QUANT and isinstance(self.mlp, Glm4MoeMLP):
+            i_q, i_s, _ = self.post_attention_layernorm(hidden_states, residual)
+            hidden_states = self.mlp(hidden_states, iqis=(i_q, i_s))
+        else:
             hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
             hidden_states = self.mlp(hidden_states)
         return hidden_states, residual