remove USE_FUSED_RMS_QUANT and USE_FUSED_SILU_MUL_QUANT

vllm/envs.py

@@ -233,8 +233,6 @@ if TYPE_CHECKING:
     VLLM_USE_LIGHTOP_MOE_SUM: bool = False
     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
@@ -1639,14 +1637,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_MERGE_ATTN_STATES_OPT":
         lambda: (os.environ.get("VLLM_USE_MERGE_ATTN_STATES_OPT", "True").lower() in
                  ("true", "1")),  
-    # vllm will use rmsquant fused op 
-    "USE_FUSED_RMS_QUANT": 
-    lambda: (os.getenv('USE_FUSED_RMS_QUANT', '0').lower() in
-             ("true", "1")),
-    # vllm will use silu_mul_quant fused op 
-    "USE_FUSED_SILU_MUL_QUANT": 
-    lambda: (os.getenv('USE_FUSED_SILU_MUL_QUANT', '0').lower() in
-             ("true", "1")),
     # vLLM will split prefill and decode, not mix up
     "VLLM_USE_PD_SPLIT":
         lambda: (os.environ.get("VLLM_USE_PD_SPLIT", "False").lower() in

vllm/model_executor/layers/linear.py

@@ -35,18 +35,6 @@ from vllm.utils import GiB_bytes
 import os
 from vllm.model_executor.utils import gemm_bank_conf
 
-if envs.USE_FUSED_RMS_QUANT:
-    try:
-        from lmslim.quantize.quant_ops import lm_faster_rmsquant
-    except Exception as e:
-        print(f"Error: Import fused rmsquant error: {e}") 
-        
-if envs.USE_FUSED_SILU_MUL_QUANT:        
-    try:
-        # from lightop import fuse_silu_mul_quant
-        from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
-    except Exception as e:
-        print(f"Error: Import fused silu_mul_qunat error: {e}")
 
 logger = init_logger(__name__)
 
@@ -441,51 +429,15 @@ class ReplicatedLinear(LinearBase):
         param.data.copy_(loaded_weight)
 
     def forward(
-
         self,
-        input_: torch.Tensor,
-        rms_weight: Optional[torch.Tensor] = None,
-        residual: Optional[torch.Tensor] = None,
-        quant_args: Optional[list] = None,
-        update_hd: Optional[bool] = True
+        x: torch.Tensor,
     ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
-        if envs.USE_FUSED_RMS_QUANT and (rms_weight is not None or quant_args is not None):
-            if quant_args is not None:
-                input_quant_args = quant_args
-            
         bias = self.bias if not self.skip_bias_add else None
         assert self.quant_method is not None
-                output = self.quant_method.apply(self, input_, bias, input_quant_args)
-                output_bias = self.bias if self.skip_bias_add else None
-                if not self.return_bias:
-                    return output
-                return output, output_bias
-
-            else:
-                i_q, _scales = lm_faster_rmsquant(input=input_,
-                                                  rms_weight=rms_weight,
-                                                  epsilon=self.eps,
-                                                  quant_dtype=torch.int8,
-                                                  residual=residual,
-                                                  update_input=update_hd
-                                                  )
 
-                new_residual = residual
-                input_quant_args = [i_q, _scales]
-                
-                bias = self.bias if not self.skip_bias_add else None
-                assert self.quant_method is not None
-                output = self.quant_method.apply(self, input_, bias, input_quant_args)
+        output = self.quant_method.apply(self, x, bias)
         output_bias = self.bias if self.skip_bias_add else None
-                if not self.return_bias:
-                    return output
-                return output, new_residual, output_bias, input_quant_args
 
-        else:
-            bias = self.bias if not self.skip_bias_add else None
-            assert self.quant_method is not None
-            output = self.quant_method.apply(self, input_, bias)
-            output_bias = self.bias if self.skip_bias_add else None
         if not self.return_bias:
             return output
         return output, output_bias
@@ -652,39 +604,13 @@ class ColumnParallelLinear(LinearBase):
     def forward(
         self,
         input_,
-        rms_weight: Optional[torch.Tensor] = None,
-        residual: Optional[torch.Tensor] = None,
-        update_hd: Optional[bool] = True
     ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
-        if envs.USE_FUSED_RMS_QUANT and rms_weight is not None:
-            input_quant_args = None
-            assert rms_weight is not None 
-            i_q, _scales = lm_faster_rmsquant(input=input_,
-                                        rms_weight=rms_weight,
-                                        epsilon=self.eps,
-                                        quant_dtype=torch.int8,
-                                        residual=residual,
-                                        update_input=update_hd)
-            new_residual = residual
-            input_quant_args = [i_q, _scales]
-        
         bias = self.bias if not self.skip_bias_add else None
 
-            assert self.quant_method is not None
-            output_parallel = self.quant_method.apply(self, input_, bias, input_quant_args)
-            if self.gather_output and self.tp_size > 1:
-                output = tensor_model_parallel_all_gather(output_parallel)
-            else:
-                output = output_parallel
-            output_bias = self.bias if self.skip_bias_add else None
-            if not self.return_bias:
-                return output
-            return output, new_residual, output_bias
-        else:
-            bias = self.bias if not self.skip_bias_add else None
         # Matrix multiply.
         assert self.quant_method is not None
         output_parallel = self.quant_method.apply(self, input_, bias)
+
         if self.gather_output and self.tp_size > 1:
             # All-gather across the partitions.
             output = tensor_model_parallel_all_gather(output_parallel)
@@ -730,54 +656,6 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
                     will be treated as a "Replicated" MergedLinear.
     """
 
-    def forward(
-        self, input_,
-        rms_weight: Optional[torch.Tensor] = None,
-        residual: Optional[torch.Tensor] = None,
-        update_hd: Optional[bool] = True
-    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
-        if envs.USE_FUSED_RMS_QUANT and rms_weight is not None:
-            input_quant_args = None
-            assert residual is not None and rms_weight is not None 
-            i_q, _scales = lm_faster_rmsquant(input=input_,
-                                        rms_weight=rms_weight,
-                                        epsilon=self.eps,
-                                        quant_dtype=torch.int8,
-                                        residual=residual,
-                                        update_input=update_hd)
-            
-            new_residual = residual
-            input_quant_args = [i_q, _scales]
-            
-            
-            bias = self.bias if not self.skip_bias_add else None
-            assert self.quant_method is not None
-            output_parallel = self.quant_method.apply(self, input_, bias, input_quant_args)
-            
-            if self.gather_output:
-                # All-gather across the partitions.
-                output = tensor_model_parallel_all_gather(output_parallel)
-            else:
-                output = output_parallel
-            output_bias = self.bias if self.skip_bias_add else None
-            if not self.return_bias:
-                return output
-            return output, new_residual, output_bias
-        else: # not USE_FUSED_RMS_QUANT
-            bias = self.bias if not self.skip_bias_add else None
-
-            assert self.quant_method is not None
-            output_parallel = self.quant_method.apply(self, input_, bias)
-            if self.gather_output and self.tp_size > 1:
-                # All-gather across the partitions.
-                output = tensor_model_parallel_all_gather(output_parallel)
-            else:
-                output = output_parallel
-            output_bias = self.bias if self.skip_bias_add else None
-            if not self.return_bias:
-                return output
-            return output, output_bias
-
     def __init__(
         self,
         input_size: int,
@@ -1548,7 +1426,6 @@ class RowParallelLinear(LinearBase):
     def forward(
         self,
         input_,
-        use_fused_silu_mul_quant: Optional[bool] = False
     ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
         if self.input_is_parallel:
             input_parallel = input_
@@ -1562,15 +1439,6 @@ 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 use_fused_silu_mul_quant:
-            xq, xs = lm_fuse_silu_mul_quant(input_parallel)
-            
-            silu_quant_args = [xq, xs]
-            output_parallel = self.quant_method.apply(self,
-                                                      input_parallel,
-                                                      bias_,
-                                                      silu_quant_args=silu_quant_args)
-        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/slimquant_w4a8.py

@@ -155,15 +155,7 @@ class SlimQuantW4A8Int8LinearMethod(LinearMethodBase):
         layer: torch.nn.Module,
         x: torch.Tensor,
         bias: Optional[torch.Tensor] = None,
-        input_quant_args: Optional[list[torch.Tensor]] = None,
-        silu_quant_args: Optional[list[torch.Tensor]] = None 
     ):
-        if envs.USE_FUSED_RMS_QUANT and input_quant_args is not None:
-            assert len(input_quant_args) == 2
-            x_q, x_scale = input_quant_args
-        elif envs.USE_FUSED_SILU_MUL_QUANT and silu_quant_args is not None:
-            x_q, x_scale = silu_quant_args
-        else:
         x_q, x_scale = per_token_quant_int8(x)
 
         if self.w8a8_strategy==1:

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

@@ -468,11 +468,7 @@ 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
-    else: # not USE_FUSED_RMS_QUANT
+
     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/deepseek_v2.py

@@ -133,20 +133,7 @@ class DeepseekV2MLP(nn.Module):
                              "Only silu is supported for now.")
         self.act_fn = SiluAndMul()
 
-    def forward(self, x,
-                rms_weight: Optional[torch.Tensor] = None,
-                residual: Optional[torch.Tensor] = None,
-                update_hd: Optional[bool] = False
-                ):
-        if envs.USE_FUSED_RMS_QUANT:
-            gate_up, new_resi, _  = self.gate_up_proj(x, rms_weight, residual, update_hd=update_hd)
-            if envs.USE_FUSED_SILU_MUL_QUANT:
-                x, _ = self.down_proj(gate_up, use_fused_silu_mul_quant=True)
-            else:
-                x = self.act_fn(gate_up)
-                x, _ = self.down_proj(x)
-            return x, new_resi
-        else:
+    def forward(self, x):
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
         x, _ = self.down_proj(x)
@@ -1128,49 +1115,6 @@ class DeepseekV2DecoderLayer(nn.Module):
         hidden_states: torch.Tensor,
         residual: Optional[torch.Tensor],
     ) -> torch.Tensor:
-        if envs.USE_FUSED_RMS_QUANT:
-            # Fix residual FP16 overflow
-            residual_fix_overflow = False
-            
-            assert self.input_layernorm.has_weight is True
-            if residual is None:
-                residual = hidden_states
-                hidden_states, _ = self.self_attn(
-                    positions = positions,
-                    hidden_states = hidden_states,
-                    rms_weight = self.input_layernorm.weight.data,
-                    residual = None
-                )
-                residual_fix_overflow = True
-            else:
-                hidden_states, new_residual = self.self_attn(
-                    positions = positions,
-                    hidden_states = hidden_states,
-                    rms_weight = self.input_layernorm.weight.data,
-                    residual = residual
-                )
-                residual = new_residual
-                
-            if hidden_states.dtype == torch.float16:
-                # rmsnorm, and rmsnorm result would not affect by scale.
-                hidden_states *= 1. / self.routed_scaling_factor
-                if self.layer_idx == 0 or residual_fix_overflow:
-                    # The residual is shared by all layers, we only scale it on
-                    # first layer.
-                    residual *= 1. / self.routed_scaling_factor
-
-            hidden_states, new_resi = self.mlp(hidden_states, self.post_attention_layernorm.weight.data, residual)
-
-            if isinstance(self.mlp,
-                        DeepseekV2MLP) and hidden_states.dtype == torch.float16:
-                # Fix FP16 overflow
-                # Scaling the DeepseekV2MLP output, it is the input of
-                # input_layernorm of next decoder layer.
-                # The scaling of DeepseekV2MOE output would be done in the forward
-                # of DeepseekV2MOE
-                hidden_states *= 1. / self.routed_scaling_factor
-            return hidden_states, new_resi
-        else:
         # Self Attention
         # Fix residual FP16 overflow
         residual_fix_overflow = False