Support MLA for DeepSeek-V2 with Triton - step 1 (#905)

python/sglang/srt/layers/extend_attention.py

@@ -57,6 +57,8 @@ def _fwd_kernel(
     stride_buf_vh,
     stride_req_to_tokens_b,
     BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
     BLOCK_M: tl.constexpr,
     BLOCK_N: tl.constexpr,
     logit_cap: tl.constexpr,
@@ -75,8 +77,10 @@ def _fwd_kernel(
     cur_batch_req_idx = tl.load(B_req_idx + cur_seq)
 
     offs_d = tl.arange(0, BLOCK_DMODEL)
+    offs_dv = tl.arange(0, BLOCK_DV)
     offs_m = tl.arange(0, BLOCK_M)
     mask_m = (cur_block_m * BLOCK_M + offs_m) < cur_seq_len_extend
+
     offs_q = (
         (cur_seq_extend_start_contiguous + cur_block_m * BLOCK_M + offs_m[:, None])
         * stride_qbs
@@ -85,10 +89,20 @@ def _fwd_kernel(
     )
     q = tl.load(Q_Extend + offs_q, mask=mask_m[:, None], other=0.0)
 
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        offs_qpe = (
+            (cur_seq_extend_start_contiguous + cur_block_m * BLOCK_M + offs_m[:, None])
+            * stride_qbs
+            + cur_head * stride_qh
+            + offs_dpe[None, :]
+        )
+        qpe = tl.load(Q_Extend + offs_qpe, mask=mask_m[:, None], other=0.0)
+
     # stage1: compute scores with prefix
     offs_n = tl.arange(0, BLOCK_N)
 
-    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DV], dtype=tl.float32)
     deno = tl.zeros([BLOCK_M], dtype=tl.float32)
     e_max = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
 
@@ -110,6 +124,18 @@ def _fwd_kernel(
 
         qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
         qk += tl.dot(q, k)
+        if BLOCK_DPE > 0:
+            offs_kpe = (
+                offs_kv_loc[None, :] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_dpe[:, None]
+            )
+            kpe = tl.load(
+                K_Buffer + offs_kpe,
+                mask=mask_n[None, :],
+                other=0.0,
+            )
+            qk += tl.dot(qpe, kpe)
         qk *= sm_scale
 
         if logit_cap > 0:
@@ -125,7 +151,7 @@ def _fwd_kernel(
         offs_buf_v = (
             offs_kv_loc[:, None] * stride_buf_vbs
             + cur_kv_head * stride_buf_vh
-            + offs_d[None, :]
+            + offs_dv[None, :]
         )
         v = tl.load(V_Buffer + offs_buf_v, mask=mask_n[:, None], other=0.0)
         p = p.to(v.dtype)
@@ -150,6 +176,21 @@ def _fwd_kernel(
 
         qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
         qk += tl.dot(q, k)
+
+        if BLOCK_DPE > 0:
+            offs_kpe = (
+                (cur_seq_extend_start_contiguous + start_n + offs_n[None, :])
+                * stride_kbs
+                + cur_kv_head * stride_kh
+                + offs_dpe[:, None]
+            )
+            kpe = tl.load(
+                K_Extend + offs_kpe,
+                mask=mask_n[None, :],
+                other=0.0,
+            )
+            qk += tl.dot(qpe, kpe)
+
         qk *= sm_scale
 
         if logit_cap > 0:
@@ -169,7 +210,7 @@ def _fwd_kernel(
         offs_v = (
             (cur_seq_extend_start_contiguous + start_n + offs_n[:, None]) * stride_vbs
             + cur_kv_head * stride_vh
-            + offs_d[None, :]
+            + offs_dv[None, :]
         )
         v = tl.load(V_Extend + offs_v, mask=mask_n[:, None], other=0.0)
         p = p.to(v.dtype)
@@ -181,7 +222,7 @@ def _fwd_kernel(
         (cur_seq_extend_start_contiguous + cur_block_m * BLOCK_M + offs_m[:, None])
         * stride_obs
         + cur_head * stride_oh
-        + offs_d[None, :]
+        + offs_dv[None, :]
     )
     tl.store(O_Extend + offs_o, acc / deno[:, None], mask=mask_m[:, None])
 
@@ -217,8 +258,17 @@ def extend_attention_fwd(
         o_extend.shape[-1],
     )
 
-    assert Lq == Lk and Lk == Lv and Lv == Lo
-    assert Lq in {16, 32, 64, 128, 256}
+    assert Lq == Lk and Lv == Lo
+    assert Lq in {16, 32, 64, 128, 256, 576}
+    assert Lv in {16, 32, 64, 128, 256, 512}
+
+    if Lq == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    else:
+        BLOCK_DMODEL = Lq
+        BLOCK_DPE = 0
+    BLOCK_DV = Lv
 
     if CUDA_CAPABILITY[0] >= 8:
         BLOCK_M, BLOCK_N = (128, 128) if Lq <= 128 else (64, 64)
@@ -260,7 +310,9 @@ def extend_attention_fwd(
         v_buffer.stride(0),
         v_buffer.stride(1),
         req_to_tokens.stride(0),
-        BLOCK_DMODEL=Lq,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
+        BLOCK_DV=BLOCK_DV,
         BLOCK_M=BLOCK_M,
         BLOCK_N=BLOCK_N,
         num_warps=num_warps,

python/sglang/srt/layers/radix_attention.py

@@ -38,16 +38,22 @@ class RadixAttention(nn.Module):
         num_kv_heads: int,
         layer_id: int,
         logit_cap: int = -1,
+        v_head_dim: int = -1,
     ):
         super().__init__()
         self.tp_q_head_num = num_heads
         self.tp_k_head_num = num_kv_heads
         self.tp_v_head_num = num_kv_heads
         self.head_dim = head_dim
+        self.qk_head_dim = head_dim
+        self.v_head_dim = v_head_dim if v_head_dim != -1 else head_dim
         self.scaling = scaling
         self.layer_id = layer_id
 
-        if not global_server_args_dict.get("disable_flashinfer", False):
+        if (
+            not global_server_args_dict.get("disable_flashinfer", False)
+            and self.qk_head_dim == self.v_head_dim
+        ):
             self.extend_forward = self.extend_forward_flashinfer
             self.decode_forward = self.decode_forward_flashinfer
         else:
@@ -57,13 +63,17 @@ class RadixAttention(nn.Module):
         self.logit_cap = logit_cap if logit_cap is not None and logit_cap > 0 else 0
 
     def extend_forward_triton(self, q, k, v, input_metadata: InputMetadata):
+        if self.qk_head_dim != self.v_head_dim:
+            o = q.new_empty((q.shape[0], self.tp_q_head_num * self.v_head_dim))
+        else:
             o = torch.empty_like(q)
+
         self.store_kv_cache(k, v, input_metadata)
         extend_attention_fwd(
-            q.view(-1, self.tp_q_head_num, self.head_dim),
+            q.view(-1, self.tp_q_head_num, self.qk_head_dim),
             k.contiguous(),
             v.contiguous(),
-            o.view(-1, self.tp_q_head_num, self.head_dim),
+            o.view(-1, self.tp_q_head_num, self.v_head_dim),
             input_metadata.token_to_kv_pool.get_key_buffer(self.layer_id),
             input_metadata.token_to_kv_pool.get_value_buffer(self.layer_id),
             input_metadata.req_to_token_pool.req_to_token,
@@ -82,14 +92,17 @@ class RadixAttention(nn.Module):
         return o
 
     def decode_forward_triton(self, q, k, v, input_metadata: InputMetadata):
+        if self.qk_head_dim != self.v_head_dim:
+            o = q.new_empty((q.shape[0], self.tp_q_head_num * self.v_head_dim))
+        else:
             o = torch.empty_like(q)
         self.store_kv_cache(k, v, input_metadata)
 
         token_attention_fwd(
-            q.view(-1, self.tp_q_head_num, self.head_dim),
+            q.view(-1, self.tp_q_head_num, self.qk_head_dim),
             input_metadata.token_to_kv_pool.get_key_buffer(self.layer_id),
             input_metadata.token_to_kv_pool.get_value_buffer(self.layer_id),
-            o.view(-1, self.tp_q_head_num, self.head_dim),
+            o.view(-1, self.tp_q_head_num, self.v_head_dim),
             input_metadata.req_to_token_pool.req_to_token,
             input_metadata.req_pool_indices,
             input_metadata.triton_start_loc,
@@ -160,8 +173,8 @@ class RadixAttention(nn.Module):
         return o.view(-1, self.tp_q_head_num * self.head_dim)
 
     def forward(self, q, k, v, input_metadata: InputMetadata):
-        k = k.view(-1, self.tp_k_head_num, self.head_dim)
-        v = v.view(-1, self.tp_v_head_num, self.head_dim)
+        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)
 
         if input_metadata.forward_mode == ForwardMode.EXTEND:
             return self.extend_forward(q, k, v, input_metadata)

python/sglang/srt/layers/token_attention.py

@@ -54,6 +54,7 @@ def _fwd_kernel_stage1(
     att_stride_h,
     kv_group_num: tl.constexpr,
     BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
     BLOCK_N: tl.constexpr,
     logit_cap: tl.constexpr,
 ):
@@ -73,6 +74,10 @@ def _fwd_kernel_stage1(
 
     off_q = cur_batch * stride_qbs + cur_head * stride_qh + offs_d
 
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        off_qpe = cur_batch * stride_qbs + cur_head * stride_qh + offs_dpe
+
     offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
 
     block_stard_index = start_n * BLOCK_N
@@ -97,6 +102,19 @@ def _fwd_kernel_stage1(
             other=0.0,
         ).to(REDUCE_TRITON_TYPE)
         att_value = tl.sum(q[None, :] * k, 1)
+        if BLOCK_DPE > 0:
+            qpe = tl.load(Q + off_qpe + start_mark).to(REDUCE_TRITON_TYPE)
+            offs_buf_kpe = (
+                k_loc[:, None] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_dpe[None, :]
+            )
+            kpe = tl.load(
+                K_Buffer + offs_buf_kpe,
+                mask=offs_n_new[:, None] < cur_batch_end_index,
+                other=0.0,
+            ).to(REDUCE_TRITON_TYPE)
+            att_value += tl.sum(qpe[None, :] * kpe, 1)
         att_value *= sm_scale
 
         if logit_cap > 0:
@@ -192,7 +210,14 @@ def _token_att_m_fwd(
     # shape constraints
     Lq, Lk = q.shape[-1], k_buffer.shape[-1]
     assert Lq == Lk
-    assert Lk in {16, 32, 64, 128, 256}
+    assert Lk in {16, 32, 64, 128, 256, 576}
+
+    if Lk == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    else:
+        BLOCK_DMODEL = Lk
+        BLOCK_DPE = 0
 
     batch, head_num = B_req_idx.shape[0], q.shape[1]
 
@@ -220,7 +245,8 @@ def _token_att_m_fwd(
         k_buffer.stride(1),
         att_out.stride(0),
         kv_group_num=kv_group_num,
-        BLOCK_DMODEL=Lk,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
         BLOCK_N=BLOCK,
         logit_cap=logit_cap,
         num_warps=num_warps,

python/sglang/srt/managers/schedule_batch.py

@@ -29,7 +29,7 @@ from sglang.global_config import global_config
 from sglang.srt.constrained import RegexGuide
 from sglang.srt.constrained.jump_forward import JumpForwardMap
 from sglang.srt.mem_cache.chunk_cache import ChunkCache
-from sglang.srt.mem_cache.memory_pool import ReqToTokenPool, TokenToKVPool
+from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
 from sglang.srt.mem_cache.radix_cache import RadixCache
 
 INIT_INCREMENTAL_DETOKENIZATION_OFFSET = 5
@@ -39,6 +39,7 @@ global_server_args_dict = {
     "disable_flashinfer": False,
     "disable_flashinfer_sampling": False,
     "attention_reduce_in_fp32": False,
+    "enable_mla": False,
 }
 
 
@@ -289,7 +290,7 @@ class Batch:
     # Request, memory pool, and cache
     reqs: List[Req]
     req_to_token_pool: ReqToTokenPool
-    token_to_kv_pool: TokenToKVPool
+    token_to_kv_pool: BaseTokenToKVPool
     tree_cache: RadixCache
 
     # Batched arguments to model runner
@@ -780,7 +781,7 @@ class InputMetadata:
     seq_lens: torch.Tensor
     positions: torch.Tensor
     req_to_token_pool: ReqToTokenPool
-    token_to_kv_pool: TokenToKVPool
+    token_to_kv_pool: BaseTokenToKVPool
 
     # For extend
     extend_seq_lens: torch.Tensor

python/sglang/srt/mem_cache/memory_pool.py

@@ -57,32 +57,18 @@ class ReqToTokenPool:
         self.can_use_mem_size = len(self.mem_state)
 
 
-class TokenToKVPool:
+class BaseTokenToKVPool:
     """A memory pool that maps a token to its kv cache locations"""
 
     def __init__(
         self,
         size: int,
-        dtype: torch.dtype,
-        head_num: int,
-        head_dim: int,
-        layer_num: int,
     ):
         self.size = size
 
         # We also add one slot. This slot is used for writing dummy output from padded tokens.
         self.mem_state = torch.ones((self.size + 1,), dtype=torch.bool, device="cuda")
 
-        # [size, head_num, head_dim] for each layer
-        self.k_buffer = [
-            torch.empty((size + 1, head_num, head_dim), dtype=dtype, device="cuda")
-            for _ in range(layer_num)
-        ]
-        self.v_buffer = [
-            torch.empty((size + 1, head_num, head_dim), dtype=dtype, device="cuda")
-            for _ in range(layer_num)
-        ]
-
         # Prefetch buffer
         self.prefetch_buffer = torch.empty(0, device="cuda", dtype=torch.int32)
         self.prefetch_chunk_size = 512
@@ -90,15 +76,6 @@ class TokenToKVPool:
         self.can_use_mem_size = self.size
         self.clear()
 
-    def get_key_buffer(self, layer_id: int):
-        return self.k_buffer[layer_id]
-
-    def get_value_buffer(self, layer_id: int):
-        return self.v_buffer[layer_id]
-
-    def get_kv_buffer(self, layer_id: int):
-        return self.k_buffer[layer_id], self.v_buffer[layer_id]
-
     def available_size(self):
         return self.can_use_mem_size + len(self.prefetch_buffer)
 
@@ -139,3 +116,67 @@ class TokenToKVPool:
 
         # We also add one slot. This slot is used for writing dummy output from padded tokens.
         self.mem_state[0] = False
+
+
+class MHATokenToKVPool(BaseTokenToKVPool):
+
+    def __init__(
+        self,
+        size: int,
+        dtype: torch.dtype,
+        head_num: int,
+        head_dim: int,
+        layer_num: int,
+    ):
+        super().__init__(size)
+
+        # [size, head_num, head_dim] for each layer
+        self.k_buffer = [
+            torch.empty((size + 1, head_num, head_dim), dtype=dtype, device="cuda")
+            for _ in range(layer_num)
+        ]
+        self.v_buffer = [
+            torch.empty((size + 1, head_num, head_dim), dtype=dtype, device="cuda")
+            for _ in range(layer_num)
+        ]
+
+    def get_key_buffer(self, layer_id: int):
+        return self.k_buffer[layer_id]
+
+    def get_value_buffer(self, layer_id: int):
+        return self.v_buffer[layer_id]
+
+    def get_kv_buffer(self, layer_id: int):
+        return self.k_buffer[layer_id], self.v_buffer[layer_id]
+
+
+class MLATokenToKVPool(BaseTokenToKVPool):
+
+    def __init__(
+        self,
+        size: int,
+        dtype: torch.dtype,
+        kv_lora_rank: int,
+        qk_rope_head_dim: int,
+        layer_num: int,
+    ):
+        super().__init__(size)
+
+        self.kv_lora_rank = kv_lora_rank
+        self.kv_buffer = [
+            torch.empty(
+                (size + 1, 1, kv_lora_rank + qk_rope_head_dim),
+                dtype=dtype,
+                device="cuda",
+            )
+            for _ in range(layer_num)
+        ]
+
+    def get_key_buffer(self, layer_id: int):
+        return self.kv_buffer[layer_id]
+
+    def get_value_buffer(self, layer_id: int):
+        return self.kv_buffer[layer_id][..., : self.kv_lora_rank]
+
+    def get_kv_buffer(self, layer_id: int):
+        return self.get_key_buffer(layer_id), self.get_value_buffer(layer_id)

python/sglang/srt/model_config.py

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 
+from enum import IntEnum, auto
 from typing import Optional
 
 from transformers import PretrainedConfig
@@ -20,6 +21,11 @@ from transformers import PretrainedConfig
 from sglang.srt.hf_transformers_utils import get_config, get_context_length
 
 
+class AttentionArch(IntEnum):
+    MLA = auto()
+    MHA = auto()
+
+
 class ModelConfig:
     def __init__(
         self,
@@ -55,6 +61,11 @@ class ModelConfig:
         # FIXME: temporary special judge for deepseek v2 MLA architecture
         if "DeepseekV2ForCausalLM" in self.hf_config.architectures:
             self.head_dim = 256
+            self.attention_arch = AttentionArch.MLA
+            self.kv_lora_rank = self.hf_config.kv_lora_rank
+            self.qk_rope_head_dim = self.hf_config.qk_rope_head_dim
+        else:
+            self.attention_arch = AttentionArch.MHA
 
         self.num_attention_heads = self.hf_config.num_attention_heads
         self.num_key_value_heads = getattr(self.hf_config, "num_key_value_heads", None)

python/sglang/srt/model_executor/model_runner.py

@@ -47,7 +47,12 @@ from sglang.srt.managers.schedule_batch import (
     InputMetadata,
     global_server_args_dict,
 )
-from sglang.srt.mem_cache.memory_pool import ReqToTokenPool, TokenToKVPool
+from sglang.srt.mem_cache.memory_pool import (
+    MHATokenToKVPool,
+    MLATokenToKVPool,
+    ReqToTokenPool,
+)
+from sglang.srt.model_config import AttentionArch
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import (
     get_available_gpu_memory,
@@ -86,6 +91,7 @@ class ModelRunner:
                 "disable_flashinfer": server_args.disable_flashinfer,
                 "disable_flashinfer_sampling": server_args.disable_flashinfer_sampling,
                 "attention_reduce_in_fp32": server_args.attention_reduce_in_fp32,
+                "enable_mla": server_args.enable_mla,
             }
         )
 
@@ -193,11 +199,19 @@ class ModelRunner:
         available_gpu_memory = get_available_gpu_memory(
             self.gpu_id, distributed=self.tp_size > 1
         )
-        head_dim = self.model_config.head_dim
-        head_num = self.model_config.get_num_kv_heads(self.tp_size)
+        if (
+            self.model_config.attention_arch == AttentionArch.MLA
+            and self.server_args.enable_mla
+        ):
+            cell_size = (
+                (self.model_config.kv_lora_rank + self.model_config.qk_rope_head_dim)
+                * self.model_config.num_hidden_layers
+                * torch._utils._element_size(self.dtype)
+            )
+        else:
             cell_size = (
-            head_num
-            * head_dim
+                self.model_config.get_num_kv_heads(self.tp_size)
+                * self.model_config.head_dim
                 * self.model_config.num_hidden_layers
                 * 2
                 * torch._utils._element_size(self.dtype)
@@ -241,7 +255,22 @@ class ModelRunner:
             max_num_reqs,
             self.model_config.context_len + 8,
         )
-        self.token_to_kv_pool = TokenToKVPool(
+        if (
+            self.model_config.attention_arch == AttentionArch.MLA
+            and self.server_args.enable_mla
+        ):
+            self.token_to_kv_pool = MLATokenToKVPool(
+                self.max_total_num_tokens,
+                dtype=self.dtype,
+                kv_lora_rank=self.model_config.kv_lora_rank,
+                qk_rope_head_dim=self.model_config.qk_rope_head_dim,
+                layer_num=self.model_config.num_hidden_layers,
+            )
+            logger.info("using MLA Triton implementaion, flashinfer is disabled")
+            # FIXME: temporarily only Triton MLA is supported
+            self.server_args.disable_flashinfer = True
+        else:
+            self.token_to_kv_pool = MHATokenToKVPool(
                 self.max_total_num_tokens,
                 dtype=self.dtype,
                 head_num=self.model_config.get_num_kv_heads(self.tp_size),

python/sglang/srt/models/deepseek_v2.py

@@ -45,6 +45,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.model_runner import InputMetadata
 
 
@@ -312,6 +313,165 @@ class DeepseekV2Attention(nn.Module):
         return output
 
 
+class DeepseekV2AttentionMLA(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        hidden_size: int,
+        num_heads: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        v_head_dim: int,
+        q_lora_rank: int,
+        kv_lora_rank: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        layer_id=None,
+    ) -> None:
+        super().__init__()
+        self.layer_id = layer_id
+        self.hidden_size = hidden_size
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.num_heads = num_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        assert num_heads % tp_size == 0
+        self.num_local_heads = num_heads // tp_size
+        self.scaling = self.qk_head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        if self.q_lora_rank is not None:
+            self.q_a_proj = ReplicatedLinear(
+                self.hidden_size,
+                self.q_lora_rank,
+                bias=False,
+                quant_config=quant_config,
+            )
+            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
+            self.q_b_proj = ColumnParallelLinear(
+                q_lora_rank,
+                self.num_heads * self.qk_head_dim,
+                bias=False,
+                quant_config=quant_config,
+            )
+        else:
+            self.q_proj = ColumnParallelLinear(
+                self.hidden_size,
+                self.num_heads * self.qk_head_dim,
+                bias=False,
+                quant_config=quant_config,
+            )
+
+        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,
+        )
+        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
+        self.kv_b_proj = ColumnParallelLinear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+            quant_config=quant_config,
+        )
+        # O projection.
+        self.o_proj = RowParallelLinear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        rope_scaling["type"] = "deepseek_yarn"
+        self.rotary_emb = get_rope(
+            qk_rope_head_dim,
+            rotary_dim=qk_rope_head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            is_neox_style=False,
+        )
+
+        if rope_scaling:
+            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
+            scaling_factor = rope_scaling["factor"]
+            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
+            self.scaling = self.scaling * mscale * mscale
+
+        self.attn = RadixAttention(
+            self.num_local_heads,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            self.scaling,
+            num_kv_heads=1,
+            layer_id=layer_id,
+            v_head_dim=self.kv_lora_rank,
+        )
+
+        kv_b_proj = self.kv_b_proj
+        w_kc, w_vc = kv_b_proj.weight.unflatten(
+            0, (-1, qk_nope_head_dim + v_head_dim)
+        ).split([qk_nope_head_dim, v_head_dim], dim=1)
+        self.w_kc = w_kc
+        self.w_vc = w_vc
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        q_len = hidden_states.shape[0]
+        q_input = hidden_states.new_empty(
+            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 = 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_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+        q_nope_out = q_input[..., : self.kv_lora_rank]
+        torch.bmm(q_nope.transpose(0, 1), self.w_kc, out=q_nope_out.transpose(0, 1))
+
+        k_input = self.kv_a_proj_with_mqa(hidden_states)[0].unsqueeze(1)
+        k_pe = k_input[..., self.kv_lora_rank :]
+        v_input = k_input[..., : self.kv_lora_rank]
+        v_input = self.kv_a_layernorm(v_input.contiguous())
+        k_input[..., : self.kv_lora_rank] = v_input
+
+        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
+        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, input_metadata)
+        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
+        attn_bmm_output = attn_output.new_empty(
+            q_len, self.num_local_heads, self.v_head_dim
+        )
+        torch.bmm(
+            attn_output.transpose(0, 1),
+            self.w_vc.transpose(1, 2).contiguous(),
+            out=attn_bmm_output.transpose(0, 1),
+        )
+
+        attn_output = attn_bmm_output.flatten(1, 2)
+        output, _ = self.o_proj(attn_output)
+
+        return output
+
+
 class DeepseekV2DecoderLayer(nn.Module):
 
     def __init__(
@@ -326,6 +486,26 @@ class DeepseekV2DecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        if global_server_args_dict["enable_mla"]:
+            self.self_attn = DeepseekV2AttentionMLA(
+                config=config,
+                hidden_size=self.hidden_size,
+                num_heads=config.num_attention_heads,
+                qk_nope_head_dim=config.qk_nope_head_dim,
+                qk_rope_head_dim=config.qk_rope_head_dim,
+                v_head_dim=config.v_head_dim,
+                q_lora_rank=(
+                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
+                ),
+                kv_lora_rank=config.kv_lora_rank,
+                rope_theta=rope_theta,
+                rope_scaling=rope_scaling,
+                max_position_embeddings=max_position_embeddings,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                layer_id=layer_id,
+            )
+        else:
             self.self_attn = DeepseekV2Attention(
                 config=config,
                 hidden_size=self.hidden_size,
@@ -333,7 +513,9 @@ class DeepseekV2DecoderLayer(nn.Module):
                 qk_nope_head_dim=config.qk_nope_head_dim,
                 qk_rope_head_dim=config.qk_rope_head_dim,
                 v_head_dim=config.v_head_dim,
-            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
+                q_lora_rank=(
+                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
+                ),
                 kv_lora_rank=config.kv_lora_rank,
                 rope_theta=rope_theta,
                 rope_scaling=rope_scaling,

python/sglang/srt/server_args.py

@@ -80,6 +80,7 @@ class ServerArgs:
     disable_disk_cache: bool = False
     enable_torch_compile: bool = False
     enable_p2p_check: bool = False
+    enable_mla: bool = False
     attention_reduce_in_fp32: bool = False
     efficient_weight_load: bool = False
 
@@ -393,6 +394,11 @@ class ServerArgs:
             action="store_true",
             help="Enable P2P check for GPU access, otherwise the p2p access is allowed by default.",
         )
+        parser.add_argument(
+            "--enable-mla",
+            action="store_true",
+            help="Enable Multi-head Latent Attention (MLA) for DeepSeek-V2",
+        )
         parser.add_argument(
             "--attention-reduce-in-fp32",
             action="store_true",