Merge branch 'v0.15.1-dev_GLM4.7_moe_call_RQ' into 'v0.15.1-dev'

perf: GLM4.7增加MOE调用rmsQuant, fix: 修掉fused_moe向后传递None导致的报错

See merge request dcutoolkit/deeplearing/vllm!505

vllm/model_executor/layers/fused_moe/layer.py

@@ -1721,7 +1721,7 @@ class FusedMoE(CustomOp):
                     hidden_states, router_logits
                 )
             else:
-                if envs.USE_FUSED_RMS_QUANT:
+                if envs.USE_FUSED_RMS_QUANT and i_q is not None and i_s is not None:
                     shared_output, fused_output = torch.ops.vllm.moe_forward_shared(
                         hidden_states, router_logits, encode_layer_name(),
                         i_q=i_q,
@@ -1749,7 +1749,10 @@ class FusedMoE(CustomOp):
         i_q: torch.Tensor | None = None,
         i_s: torch.Tensor | None = None
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
-        return self.forward_native(hidden_states, router_logits, i_q=i_q, i_s=i_s)
+        if envs.USE_FUSED_RMS_QUANT and i_q is not None and i_s is not None:
+            return self.forward_native(hidden_states, router_logits, i_q=i_q, i_s=i_s)
+        else:
+            return self.forward_native(hidden_states, router_logits)
 
     def forward_impl_chunked(
         self,
@@ -1981,7 +1984,7 @@ class FusedMoE(CustomOp):
             # because matrix multiply maybe modify the hidden_states.
             if has_separate_shared_experts and not use_shared_experts_stream:
                 assert self.shared_experts is not None
-                if envs.USE_FUSED_RMS_QUANT:
+                if envs.USE_FUSED_RMS_QUANT and i_q is not None and i_s is not None:
                     shared_output = self.shared_experts(hidden_states, iqis=(i_q, i_s))
                 else:
                     shared_output = self.shared_experts(hidden_states)
@@ -2196,7 +2199,7 @@ def moe_forward_shared(
 ) -> tuple[torch.Tensor, torch.Tensor]:
     self = get_layer_from_name(layer_name)
     assert self.shared_experts is not None
-    if envs.USE_FUSED_RMS_QUANT:
+    if envs.USE_FUSED_RMS_QUANT and i_q is not None and i_s is not None:
         return self.forward_impl(hidden_states, router_logits, i_q=i_q, i_s=i_s)
     else:
         return self.forward_impl(hidden_states, router_logits)

vllm/model_executor/layers/fused_moe/shared_fused_moe.py

@@ -8,6 +8,7 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.fused_moe.layer import FusedMoE
+from vllm import envs
 
 
 # TODO(bnell): Add shared + fused combo function? e.g. +
@@ -64,9 +65,14 @@ class SharedFusedMoE(FusedMoE):
     ) -> tuple[torch.Tensor, torch.Tensor]:
         if not self.use_overlapped:
             if self._shared_experts is not None:
-                shared_out = self._shared_experts(hidden_states,
-                i_q=iqis[0] if iqis is not None else None,
-                i_s=iqis[1] if iqis is not None else None)
+                if envs.USE_FUSED_RMS_QUANT and iqis is not None:
+                    assert iqis[0] is not None
+                    assert iqis[1] is not None
+                    shared_out = self._shared_experts(hidden_states,
+                    i_q=iqis[0],
+                    i_s=iqis[1])
+                else:
+                    shared_out = self._shared_experts(hidden_states)
 
                 # Reduce shared expert outputs if necessary, since the MLP
                 # should have been created with reduce_results=False.
@@ -79,19 +85,35 @@ class SharedFusedMoE(FusedMoE):
             else:
                 shared_out = None
 
-            fused_out = super().forward(
-                hidden_states=hidden_states,
-                router_logits=router_logits,
-                i_q=iqis[0] if iqis is not None else None,
-                i_s=iqis[1] if iqis is not None else None,
-            )
+            if envs.USE_FUSED_RMS_QUANT and iqis is not None:
+                assert iqis[0] is not None
+                assert iqis[1] is not None
+                fused_out = super().forward(
+                    hidden_states=hidden_states,
+                    router_logits=router_logits,
+                    i_q=iqis[0],
+                    i_s=iqis[1]
+                )
+            else:
+                fused_out = super().forward(
+                    hidden_states=hidden_states,
+                    router_logits=router_logits
+                )
         else:
-            shared_out, fused_out = super().forward(
-                hidden_states=hidden_states,
-                router_logits=router_logits,
-                i_q=iqis[0] if iqis is not None else None,
-                i_s=iqis[1] if iqis is not None else None,
-            )
+            if envs.USE_FUSED_RMS_QUANT and iqis is not None:
+                assert iqis[0] is not None
+                assert iqis[1] is not None
+                shared_out, fused_out = super().forward(
+                    hidden_states=hidden_states,
+                    router_logits=router_logits,
+                    i_q=iqis[0],
+                    i_s=iqis[1]
+                )
+            else:
+                shared_out, fused_out = super().forward(
+                    hidden_states=hidden_states, 
+                    router_logits=router_logits
+                )
             # ensure early TP reduction of shared expert outputs when required
             if (
                 shared_out is not None

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

@@ -170,10 +170,14 @@ def apply_int8_linear(
     # * 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
+        assert input_quant_args[0] is not None
+        assert input_quant_args[1] is not None
         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
+        assert silu_quant_args[0] is not None
+        assert silu_quant_args[1] is not None
         x_zp =None
         x_q, x_scale = silu_quant_args
     else:

vllm/model_executor/models/glm4_moe.py

@@ -110,12 +110,13 @@ class Glm4MoeMLP(nn.Module):
 
     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:
+        if envs.USE_FUSED_SILU_MUL_QUANT and iqis is not None:
+            gate_up, _ = self.gate_up_proj(x, iqis=iqis)
             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:
+            gate_up, _ = self.gate_up_proj(x)
             x = self.act_fn(gate_up)
             x, _ = self.down_proj(x)
         return x
@@ -207,16 +208,22 @@ class Glm4MoE(nn.Module):
             router_logits_dtype=torch.float32,
         )
 
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    def forward(self, 
+                hidden_states: torch.Tensor,
+                iqis: tuple[torch.Tensor, torch.Tensor] | None = None
+                ) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
 
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states.to(dtype=torch.float32))
-
-        fused_moe_out = self.experts(
-            hidden_states=hidden_states, router_logits=router_logits
-        )
+        if envs.USE_FUSED_RMS_QUANT and iqis is not None:
+            fused_moe_out = self.experts(
+                hidden_states=hidden_states, router_logits=router_logits,
+                iqis=iqis)
+        else:
+            fused_moe_out = self.experts(
+            hidden_states=hidden_states, router_logits=router_logits)
 
         if self.shared_experts is not None:
             shared_output, final_hidden_states = fused_moe_out
@@ -420,7 +427,7 @@ class Glm4MoeDecoderLayer(nn.Module):
         else:
             self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-        if envs.USE_FUSED_RMS_QUANT and isinstance(self.mlp, Glm4MoeMLP):
+        if envs.USE_FUSED_RMS_QUANT:
             self.post_attention_layernorm = FusedRMSNormQuant(
                 config.hidden_size, eps=config.rms_norm_eps
             )
@@ -459,8 +466,12 @@ class Glm4MoeDecoderLayer(nn.Module):
                                                       )
             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)
+        if envs.USE_FUSED_RMS_QUANT:
+            update_hs = True if isinstance(self.mlp, Glm4MoE) else False
+            i_q, i_s, residual = self.post_attention_layernorm(x=hidden_states,
+                                                               residual=residual,
+                                                               quant_dtype=torch.int8,
+                                                               update_input=update_hs)
             hidden_states = self.mlp(hidden_states, iqis=(i_q, i_s))
         else:
             hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)