Fuse q_a_proj and kv_a_proj (#5619)

python/sglang/srt/models/deepseek_v2.py

@@ -443,12 +443,12 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         # For tensor parallel attention
         if self.q_lora_rank is not None:
-            self.q_a_proj = ReplicatedLinear(
+            self.fused_qkv_a_proj_with_mqa = ReplicatedLinear(
                 self.hidden_size,
-                self.q_lora_rank,
+                self.q_lora_rank + self.kv_lora_rank + self.qk_rope_head_dim,
                 bias=False,
                 quant_config=quant_config,
-                prefix=add_prefix("q_a_proj", prefix),
+                prefix=add_prefix("fused_qkv_a_proj_with_mqa", prefix),
             )
             self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
             self.q_b_proj = ColumnParallelLinear(
@@ -470,6 +470,14 @@ class DeepseekV2AttentionMLA(nn.Module):
                 tp_rank=attn_tp_rank,
                 tp_size=attn_tp_size,
             )
+            self.kv_a_proj_with_mqa = ReplicatedLinear(
+                self.hidden_size,
+                self.kv_lora_rank + self.qk_rope_head_dim,
+                bias=False,
+                quant_config=quant_config,
+                prefix=add_prefix("kv_a_proj_with_mqa", prefix),
+            )
+
         self.kv_b_proj = ColumnParallelLinear(
             self.kv_lora_rank,
             self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
@@ -490,14 +498,6 @@ class DeepseekV2AttentionMLA(nn.Module):
             tp_rank=attn_tp_rank,
             tp_size=attn_tp_size,
         )
-
-        self.kv_a_proj_with_mqa = ReplicatedLinear(
-            self.hidden_size,
-            self.kv_lora_rank + self.qk_rope_head_dim,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_a_proj_with_mqa", prefix),
-        )
         self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
 
         if rope_scaling:
@@ -656,15 +656,18 @@ class DeepseekV2AttentionMLA(nn.Module):
         forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
+            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
+                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
+            )
             q = self.q_a_layernorm(q)
             q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
         else:
             q = self.q_proj(hidden_states)[0].view(
                 -1, self.num_local_heads, self.qk_head_dim
             )
-        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
             latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
+
+        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
         kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
         latent_cache = latent_cache.unsqueeze(1)
         kv_a = self.kv_a_layernorm(kv_a.contiguous())
@@ -699,13 +702,16 @@ class DeepseekV2AttentionMLA(nn.Module):
         zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
         if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
+            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
+                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
+            )
             q = self.q_a_layernorm(q)
             q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
         else:
             q = self.q_proj(hidden_states)[0].view(
                 -1, self.num_local_heads, self.qk_head_dim
             )
+            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
 
         if self.use_deep_gemm_bmm:
@@ -744,7 +750,6 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         q_nope_out = q_nope_out.transpose(0, 1)
 
-        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         k_nope = latent_cache[..., : self.kv_lora_rank]
         k_nope = self.kv_a_layernorm(k_nope).unsqueeze(1)
         k_pe = latent_cache[..., self.kv_lora_rank :].unsqueeze(1)
@@ -819,13 +824,16 @@ class DeepseekV2AttentionMLA(nn.Module):
             q_len, self.num_local_heads, self.kv_lora_rank + self.qk_rope_head_dim
         )
         if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
+            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
+                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
+            )
             q = self.q_a_layernorm(q)
             q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
         else:
             q = self.q_proj(hidden_states)[0].view(
                 -1, self.num_local_heads, self.qk_head_dim
             )
+            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
 
         if self.w_kc.dtype == torch.float8_e4m3fnuz:
@@ -846,8 +854,6 @@ class DeepseekV2AttentionMLA(nn.Module):
         else:
             q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
         q_input[..., : self.kv_lora_rank] = q_nope_out.transpose(0, 1)
-
-        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         v_input = latent_cache[..., : self.kv_lora_rank]
         v_input = self.kv_a_layernorm(v_input.contiguous()).unsqueeze(1)
         k_input = latent_cache.unsqueeze(1)
@@ -1018,15 +1024,17 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         # First do normal mha forward to get output for extended part
         if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
+            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
+                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
+            )
             q = self.q_a_layernorm(q)
             q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
         else:
             q = self.q_proj(hidden_states)[0].view(
                 -1, self.num_local_heads, self.qk_head_dim
             )
-        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
             latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
+        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
         kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
         latent_cache = latent_cache.unsqueeze(1)
         kv_a = self.kv_a_layernorm(kv_a.contiguous())
@@ -1668,6 +1676,12 @@ class DeepseekV2ForCausalLM(nn.Module):
             num_experts=self.config.n_routed_experts + self.n_share_experts_fusion,
         )
 
+        # Fuse q_a_proj and kv_a_proj_with_mqa along output dimension when q_lora_rank is not None
+        fuse_qkv_a_proj = hasattr(self.config, "q_lora_rank") and (
+            self.config.q_lora_rank is not None
+        )
+        cached_a_proj = {} if fuse_qkv_a_proj else None
+
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
             # TODO(HandH1998): Modify it when nextn is supported.
@@ -1723,6 +1737,45 @@ class DeepseekV2ForCausalLM(nn.Module):
                     if name.endswith(".bias") and name not in params_dict:
                         continue
 
+                    if fuse_qkv_a_proj and (
+                        "q_a_proj" in name or "kv_a_proj_with_mqa" in name
+                    ):
+                        cached_a_proj[name] = loaded_weight
+                        q_a_proj_name = (
+                            name
+                            if "q_a_proj" in name
+                            else name.replace("kv_a_proj_with_mqa", "q_a_proj")
+                        )
+                        kv_a_proj_name = (
+                            name
+                            if "kv_a_proj_with_mqa" in name
+                            else name.replace("q_a_proj", "kv_a_proj_with_mqa")
+                        )
+
+                        # When both q_a_proj and kv_a_proj_with_mqa has been cached, load the fused weight to parameter
+                        if (
+                            q_a_proj_name in cached_a_proj
+                            and kv_a_proj_name in cached_a_proj
+                        ):
+
+                            q_a_proj_weight = cached_a_proj[q_a_proj_name]
+                            kv_a_proj_weight = cached_a_proj[kv_a_proj_name]
+                            fused_weight = torch.cat(
+                                [q_a_proj_weight, kv_a_proj_weight], dim=0
+                            )
+
+                            param_name = name.replace(
+                                "q_a_proj", "fused_qkv_a_proj_with_mqa"
+                            )
+                            param = params_dict[param_name]
+
+                            weight_loader = getattr(
+                                param, "weight_loader", default_weight_loader
+                            )
+                            weight_loader(param, fused_weight)
+                            cached_a_proj.pop(q_a_proj_name)
+                            cached_a_proj.pop(kv_a_proj_name)
+                    else:
                         param = params_dict[name]
                         weight_loader = getattr(
                             param, "weight_loader", default_weight_loader