sglang 0.5.4__py3-none-any.whl → 0.5.4.post1__py3-none-any.whl

sglang/srt/models/deepseek_v2.py

@@ -57,6 +57,7 @@ from sglang.srt.layers.attention.npu_ops.mla_preprocess import (
     is_mla_preprocess_enabled,
 )
 from sglang.srt.layers.attention.nsa.nsa_indexer import Indexer
+from sglang.srt.layers.attention.utils import concat_and_cast_mha_k_triton
 from sglang.srt.layers.communicator import (
     LayerCommunicator,
     LayerScatterModes,
@@ -241,6 +242,10 @@ class AttnForwardMethod(IntEnum):
     # This method can avoid OOM when prefix lengths are long.
     MHA_CHUNKED_KV = auto()
 
+    # Use multi-head attention, execute the MHA for prefix and extended kv in one shot
+    # when the sequence lengths are below the threshold.
+    MHA_ONE_SHOT = auto()
+
     # Use MLA but with fused RoPE
     MLA_FUSED_ROPE = auto()
 
@@ -306,6 +311,14 @@ def _is_extend_without_speculative(forward_batch):
     )
 
 
+def _support_mha_one_shot(attn: DeepseekV2AttentionMLA, forward_batch, backend_name):
+    attn_supported = backend_name in ["fa3", "flashinfer", "flashmla"]
+    sum_seq_lens = (
+        sum(forward_batch.seq_lens_cpu) if forward_batch.seq_lens_cpu is not None else 0
+    )
+    return attn_supported and sum_seq_lens <= forward_batch.get_max_chunk_capacity()
+
+
 def _handle_attention_backend(
     attn: DeepseekV2AttentionMLA, forward_batch, backend_name
 ):
@@ -325,6 +338,8 @@ def _handle_attention_backend(
             or sum_extend_prefix_lens == 0
         )
     ):
+        if _support_mha_one_shot(attn, forward_batch, backend_name):
+            return AttnForwardMethod.MHA_ONE_SHOT
         return AttnForwardMethod.MHA_CHUNKED_KV
     else:
         return _dispatch_mla_subtype(attn, forward_batch)
@@ -335,7 +350,11 @@ def handle_attention_flashinfer(attn, forward_batch):
 
 
 def handle_attention_fa3(attn, forward_batch):
-    return _handle_attention_backend(attn, forward_batch, "fa3")
+    # when deterministic inference is enabled, use MLA
+    if get_global_server_args().enable_deterministic_inference:
+        return _dispatch_mla_subtype(attn, forward_batch)
+    else:
+        return _handle_attention_backend(attn, forward_batch, "fa3")
 
 
 def handle_attention_flashmla(attn, forward_batch):
@@ -379,6 +398,10 @@ def handle_attention_nsa(attn, forward_batch):
 
 
 def handle_attention_triton(attn, forward_batch):
+    # when deterministic inference is enabled, use MLA
+    if get_global_server_args().enable_deterministic_inference:
+        return _dispatch_mla_subtype(attn, forward_batch)
+
     if (
         _is_extend_without_speculative(forward_batch)
         and sum(forward_batch.extend_prefix_lens_cpu) == 0
@@ -982,16 +1005,14 @@ class DeepseekV2MoE(nn.Module):
                 )
 
     def op_experts(self, state):
-        state.hidden_states_experts_output = self.experts.run_moe_core(
+        state.combine_input = self.experts.run_moe_core(
             dispatch_output=state.dispatch_output,
         )
 
     def op_combine_a(self, state):
         if self.ep_size > 1:
             self.experts.dispatcher.combine_a(
-                hidden_states=state.pop("hidden_states_experts_output"),
-                topk_ids=state.dispatch_output.topk_ids,
-                topk_weights=state.dispatch_output.topk_weights,
+                combine_input=state.pop("combine_input"),
                 tbo_subbatch_index=state.get("tbo_subbatch_index"),
             )
             state.pop("dispatch_output")
@@ -1062,6 +1083,7 @@ class DeepseekV2AttentionMLA(nn.Module):
         self.scaling = self.qk_head_dim**-0.5
         self.rope_theta = rope_theta
         self.max_position_embeddings = max_position_embeddings
+        self.kv_cache_dtype = get_global_server_args().kv_cache_dtype
 
         # NOTE modification to rope_scaling must be done early enough, b/c e.g. Indexer needs it
         if rope_scaling:
@@ -1359,6 +1381,10 @@ class DeepseekV2AttentionMLA(nn.Module):
             inner_state = self.forward_normal_chunked_kv_prepare(
                 positions, hidden_states, forward_batch, zero_allocator
             )
+        elif attn_forward_method == AttnForwardMethod.MHA_ONE_SHOT:
+            inner_state = self.forward_normal_one_shot_prepare(
+                positions, hidden_states, forward_batch, zero_allocator
+            )
         elif attn_forward_method == AttnForwardMethod.MLA:
             if not self.is_mla_preprocess_enabled:
                 inner_state = self.forward_absorb_prepare(
@@ -1410,6 +1436,8 @@ class DeepseekV2AttentionMLA(nn.Module):
             return self.forward_normal_core(*inner_state)
         elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
             return self.forward_normal_chunked_kv_core(*inner_state)
+        elif attn_forward_method == AttnForwardMethod.MHA_ONE_SHOT:
+            return self.forward_normal_one_shot_core(*inner_state)
         elif attn_forward_method == AttnForwardMethod.MLA:
             return self.forward_absorb_core(*inner_state)
         elif attn_forward_method == AttnForwardMethod.NPU_MLA_SPARSE:
@@ -1444,41 +1472,24 @@ class DeepseekV2AttentionMLA(nn.Module):
         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)
-        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)
 
-        # Temporary for DeepSeek V3/R1 only, but can generalize if needed
+        self._set_mla_kv_buffer(latent_cache, kv_a, k_pe, forward_batch)
         if (
-            _is_cuda
-            and (self.num_local_heads == 128)
-            and (self.qk_nope_head_dim == 128)
-            and (self.qk_rope_head_dim == 64)
+            forward_batch.mha_one_shot
+            and sum(forward_batch.extend_prefix_lens_cpu) != 0
         ):
-            concat_mla_k(k=k, k_nope=k_nope, k_rope=k_pe)
-        else:
-            k[..., : self.qk_nope_head_dim] = k_nope
-            k[..., self.qk_nope_head_dim :] = k_pe
-
-        if not _is_npu:
-            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
-            )
-        else:
-            # To reduce a time-costing split operation
-            forward_batch.token_to_kv_pool.set_kv_buffer(
-                self.attn_mha, forward_batch.out_cache_loc, kv_a.unsqueeze(1), k_pe
+            kv_a, k_pe = self._get_mla_kv_buffer(
+                forward_batch.fetch_mha_one_shot_kv_indices(), q.dtype, forward_batch
             )
+        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 = self._concat_and_cast_mha_k(k_nope, k_pe, forward_batch)
         return q, k, v, forward_batch
 
     def forward_normal_core(self, q, k, v, forward_batch):
@@ -2288,20 +2299,11 @@ class DeepseekV2AttentionMLA(nn.Module):
         for i in range(forward_batch.num_prefix_chunks):
             forward_batch.set_prefix_chunk_idx(i)
 
+            kv_indices = forward_batch.prefix_chunk_kv_indices[i]
             # Fetch latent cache from memory pool with precomputed chunked kv indices
-            latent_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
-                self.attn_mha.layer_id
-            )
-            latent_cache = (
-                latent_cache_buf[forward_batch.prefix_chunk_kv_indices[i]]
-                .contiguous()
-                .to(q.dtype)
+            kv_a_normed, k_pe = self._get_mla_kv_buffer(
+                kv_indices, q.dtype, forward_batch
             )
-
-            kv_a_normed, k_pe = latent_cache.split(
-                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
-            )
-            kv_a_normed = kv_a_normed.squeeze(1).contiguous()
             kv = self.kv_b_proj(kv_a_normed)[0]
             kv = kv.view(
                 -1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim
@@ -2376,6 +2378,107 @@ class DeepseekV2AttentionMLA(nn.Module):
         output, _ = self.o_proj(attn_output)
         return output
 
+    def forward_normal_one_shot_prepare(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        zero_allocator: BumpAllocator,
+    ):
+        forward_batch.mha_one_shot = True
+        return self.forward_normal_prepare(
+            positions, hidden_states, forward_batch, zero_allocator
+        )
+
+    def forward_normal_one_shot_core(self, q, k, v, forward_batch):
+        has_extend_prefix = any(forward_batch.extend_prefix_lens_cpu)
+        # Only initialize the info once
+        if has_extend_prefix and forward_batch.num_prefix_chunks is None:
+            forward_batch.num_prefix_chunks = 0
+            if hasattr(forward_batch.attn_backend, "init_mha_chunk_metadata"):
+                forward_batch.attn_backend.init_mha_chunk_metadata(forward_batch)
+        forward_batch.mha_return_lse = False
+        # Do mha for extended part without prefix
+        forward_batch.set_attn_attend_prefix_cache(False)
+        return self.forward_normal_core(q, k, v, forward_batch)
+
+    def _set_mla_kv_buffer(
+        self,
+        latent_cache: torch.Tensor,
+        kv_a: torch.Tensor,
+        k_pe: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ):
+        if _is_cuda:
+            # Save latent cache
+            forward_batch.token_to_kv_pool.set_mla_kv_buffer(
+                self.attn_mha, forward_batch.out_cache_loc, kv_a.unsqueeze(1), k_pe
+            )
+        elif _is_npu:
+            # To reduce a time-costing split operation
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                self.attn_mha, forward_batch.out_cache_loc, kv_a.unsqueeze(1), k_pe
+            )
+        else:
+            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
+            )
+
+    def _get_mla_kv_buffer(
+        self,
+        kv_indices: torch.Tensor,
+        dst_dtype: torch.dtype,
+        forward_batch: ForwardBatch,
+    ):
+        if _is_cuda:
+            kv_a, k_pe = forward_batch.token_to_kv_pool.get_mla_kv_buffer(
+                self.attn_mha, kv_indices, dst_dtype
+            )
+            kv_a = kv_a.squeeze(1)
+        else:
+            latent_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
+                self.attn_mha.layer_id
+            )
+            latent_cache = latent_cache_buf[kv_indices].contiguous().to(dst_dtype)
+
+            kv_a, k_pe = latent_cache.split(
+                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+            )
+            kv_a = kv_a.squeeze(1).contiguous()
+        return kv_a, k_pe
+
+    def _concat_and_cast_mha_k(self, k_nope, k_pe, forward_batch):
+        # Temporary for DeepSeek V3/R1 only, but can generalize if needed
+        k_shape = (k_nope.shape[0], self.num_local_heads, self.qk_head_dim)
+        if (
+            _is_cuda
+            and (self.num_local_heads == 128)
+            and (self.qk_nope_head_dim == 128)
+            and (self.qk_rope_head_dim == 64)
+        ):
+            k = k_nope.new_empty(*k_shape)
+            concat_mla_k(k=k, k_nope=k_nope, k_rope=k_pe)
+        elif _is_cuda:
+            # fa3 mha support fp8 inputs
+            if (
+                self.current_attention_backend == "fa3"
+                and self.kv_cache_dtype != "auto"
+            ):
+                attn_dtype = forward_batch.token_to_kv_pool.dtype
+            else:
+                attn_dtype = k_nope.dtype
+            k = k_nope.new_empty(*k_shape, dtype=attn_dtype)
+            concat_and_cast_mha_k_triton(k, k_nope, k_pe)
+        else:
+            k = k_nope.new_empty(*k_shape)
+            k[..., : self.qk_nope_head_dim] = k_nope
+            k[..., self.qk_nope_head_dim :] = k_pe
+        return k
+
     @staticmethod
     def _get_q_b_proj_quant_config(quant_config):
         if get_bool_env_var("SGLANG_NVFP4_CKPT_FP8_GEMM_IN_ATTN"):