MLA prefill w/o weight absorption (#2349)

python/sglang/srt/layers/attention/__init__.py

@@ -52,12 +52,13 @@ def forward(
         v: torch.Tensor,
         layer: RadixAttention,
         forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
     ):
         """Run forward on an attention layer."""
         if forward_batch.forward_mode.is_decode():
-            return self.forward_decode(q, k, v, layer, forward_batch)
+            return self.forward_decode(q, k, v, layer, forward_batch, save_kv_cache)
         else:
-            return self.forward_extend(q, k, v, layer, forward_batch)
+            return self.forward_extend(q, k, v, layer, forward_batch, save_kv_cache)
 
     def forward_decode(
         self,
@@ -66,6 +67,7 @@ def forward_decode(
         v: torch.Tensor,
         layer: RadixAttention,
         forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
     ):
         """Run a forward for decode."""
         raise NotImplementedError()
@@ -77,6 +79,7 @@ def forward_extend(
         v: torch.Tensor,
         layer: RadixAttention,
         forward_batch: ForwardBatch,
+        save_kv_cache: bool = True,
     ):
         """Run a forward for extend."""
         raise NotImplementedError()

python/sglang/srt/layers/attention/double_sparsity_backend.py

@@ -165,7 +165,13 @@ def get_cuda_graph_seq_len_fill_value(self):
         return 1
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # TODO: reuse the buffer across layers
         if layer.qk_head_dim != layer.v_head_dim:
@@ -181,9 +187,10 @@ def forward_extend(
             .expand(k.shape[0], -1, -1),
         )
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v, k_label
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v, k_label
+            )
 
         (
             start_loc,
@@ -212,7 +219,13 @@ def forward_extend(
         return o
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -242,9 +255,10 @@ def forward_decode(
             .expand(k.shape[0], -1, -1),
         )
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v, k_label
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v, k_label
+            )
 
         # NOTE(Andy) shouldn't be used when max_len_in_batch < heavy_token_num
         #            and set a minimum value for sparse_decode

python/sglang/srt/layers/attention/flashinfer_backend.py

@@ -221,7 +221,13 @@ def get_cuda_graph_seq_len_fill_value(self):
         return 0
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         prefill_wrapper_paged = self.prefill_wrappers_paged[
             self._get_wrapper_idx(layer)
@@ -237,7 +243,8 @@ def forward_extend(
         if not use_ragged:
             if k is not None:
                 assert v is not None
-                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+                if save_kv_cache:
+                    forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
             o = prefill_wrapper_paged.forward(
                 q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
@@ -270,12 +277,19 @@ def forward_extend(
 
                 o, _ = merge_state(o1, s1, o2, s2)
 
-            forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+            if save_kv_cache:
+                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
         return o.view(-1, layer.tp_q_head_num * layer.head_dim)
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         decode_wrapper = self.forward_metadata[0][self._get_wrapper_idx(layer)]
         cache_loc = (
@@ -286,7 +300,8 @@ def forward_decode(
 
         if k is not None:
             assert v is not None
-            forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+            if save_kv_cache:
+                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
         o = decode_wrapper.forward(
             q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),

python/sglang/srt/layers/attention/torch_native_backend.py

@@ -216,16 +216,23 @@ def _run_sdpa_forward_decode(
         return output
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         if layer.qk_head_dim != layer.v_head_dim:
             o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
         else:
             o = torch.empty_like(q)
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 
@@ -249,7 +256,13 @@ def forward_extend(
         return o
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -260,9 +273,10 @@ def forward_decode(
         else:
             o = torch.empty_like(q)
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 

python/sglang/srt/layers/attention/triton_backend.py

@@ -114,17 +114,24 @@ def get_cuda_graph_seq_len_fill_value(self):
         return 1
 
     def forward_extend(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # TODO: reuse the buffer across layers
         if layer.qk_head_dim != layer.v_head_dim:
             o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
         else:
             o = torch.empty_like(q)
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         start_loc, attn_logits, max_seq_len, max_extend_len = self.forward_metadata
         self.extend_attention_fwd(
@@ -146,7 +153,13 @@ def forward_extend(
         return o
 
     def forward_decode(
-        self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -160,9 +173,10 @@ def forward_decode(
 
         start_loc, attn_logits, max_seq_len, max_extend_len = self.forward_metadata
 
-        forward_batch.token_to_kv_pool.set_kv_buffer(
-            layer, forward_batch.out_cache_loc, k, v
-        )
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
 
         self.decode_attention_fwd(
             q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),

python/sglang/srt/layers/attention/triton_ops/extend_attention.py

@@ -284,6 +284,9 @@ def extend_attention_fwd(
     elif Lq == 288:
         BLOCK_DMODEL = 256
         BLOCK_DPE = 32
+    elif Lq == 192:
+        BLOCK_DMODEL = 128
+        BLOCK_DPE = 64
     else:
         BLOCK_DMODEL = triton.next_power_of_2(Lq)
         BLOCK_DPE = 0

python/sglang/srt/layers/radix_attention.py

@@ -48,11 +48,13 @@ def __init__(
         self.sliding_window_size = sliding_window_size or -1
         self.is_cross_attention = is_cross_attention
 
-    def forward(self, q, k, v, forward_batch: ForwardBatch):
+    def forward(self, q, k, v, forward_batch: ForwardBatch, save_kv_cache=True):
         if k is not None:
             # For cross-layer sharing, kv can be None
             assert v is not None
             k = k.view(-1, self.tp_k_head_num, self.qk_head_dim)
             v = v.view(-1, self.tp_v_head_num, self.v_head_dim)
 
-        return forward_batch.attn_backend.forward(q, k, v, self, forward_batch)
+        return forward_batch.attn_backend.forward(
+            q, k, v, self, forward_batch, save_kv_cache
+        )

python/sglang/srt/models/deepseek_v2.py

@@ -453,7 +453,7 @@ def __init__(
             mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
             self.scaling = self.scaling * mscale * mscale
 
-        self.attn = RadixAttention(
+        self.attn_mqa = RadixAttention(
             self.num_local_heads,
             self.kv_lora_rank + self.qk_rope_head_dim,
             self.scaling,
@@ -462,6 +462,15 @@ def __init__(
             v_head_dim=self.kv_lora_rank,
         )
 
+        self.attn_mha = RadixAttention(
+            self.num_local_heads,
+            self.qk_nope_head_dim + self.qk_rope_head_dim,
+            self.scaling,
+            num_kv_heads=self.num_local_heads,
+            layer_id=layer_id,
+            v_head_dim=self.v_head_dim,
+        )
+
         self.w_kc = None
         self.w_vc = None
         self.w_scale = None
@@ -471,6 +480,63 @@ def forward(
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        # Use normal computation for prefill and use weight absorption for extend/decode
+        if (
+            forward_batch.forward_mode.is_extend()
+            and forward_batch.extend_prefix_lens.sum() == 0
+        ):
+            return self.forward_normal(positions, hidden_states, forward_batch)
+        else:
+            return self.forward_absorb(positions, hidden_states, forward_batch)
+
+    def forward_normal(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        if self.q_lora_rank is not None:
+            q = self.q_a_proj(hidden_states)[0]
+            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]
+        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())
+        kv = self.kv_b_proj(kv_a)[0]
+        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
+        k_nope = kv[..., : self.qk_nope_head_dim]
+        v = kv[..., self.qk_nope_head_dim :]
+        k_pe = latent_cache[:, :, self.kv_lora_rank :]
+        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
+        q[..., self.qk_nope_head_dim :] = q_pe
+        k = torch.empty_like(q)
+        k[..., : self.qk_nope_head_dim] = k_nope
+        k[..., self.qk_nope_head_dim :] = k_pe
+
+        latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
+        latent_cache[:, :, self.kv_lora_rank :] = k_pe
+
+        # Save latent cache
+        forward_batch.token_to_kv_pool.set_kv_buffer(
+            self.attn_mha, forward_batch.out_cache_loc, latent_cache, None
+        )
+        attn_output = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
+        attn_output = attn_output.reshape(-1, self.num_local_heads * self.v_head_dim)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+    def forward_absorb(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         q_len = hidden_states.shape[0]
         q_input = hidden_states.new_empty(
@@ -508,7 +574,7 @@ def forward(
         q_input[..., self.kv_lora_rank :] = q_pe
         k_input[..., self.kv_lora_rank :] = k_pe
 
-        attn_output = self.attn(q_input, k_input, v_input, forward_batch)
+        attn_output = self.attn_mqa(q_input, k_input, v_input, forward_batch)
         attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
 
         if self.w_vc.dtype == torch.float8_e4m3fn:
@@ -835,7 +901,6 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
                 if hasattr(self_attn.kv_b_proj, "weight_scale"):
                     self_attn.w_scale = self_attn.kv_b_proj.weight_scale
-                del self_attn.kv_b_proj
 
 
 EntryClass = DeepseekV2ForCausalLM