[New Model] DeepSeek-V3.2 (Rebased to Main) (#25896)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Signed-off-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
Signed-off-by: Lucia Fang <fanglu@meta.com>
Co-authored-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: yewentao256 <zhyanwentao@126.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Lucia Fang <116399278+luccafong@users.noreply.github.com>
Co-authored-by: Lucia Fang <fanglu@meta.com>
Co-authored-by: NickLucche <nlucches@redhat.com>
Co-authored-by: Siyuan Fu <siyuanf@nvidia.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Xiaozhu Meng <mxz297@gmail.com>
Co-authored-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
Signed-off-by: simon-mo <simon.mo@hey.com>

tests/v1/core/test_single_type_kv_cache_manager.py

@@ -35,7 +35,6 @@ def test_chunked_local_attention_possible_cached_prefix():
         head_size=1,
         dtype=torch.float32,
         attention_chunk_size=4,
-        use_mla=False,
     )
 
     block_pool = BlockPool(num_gpu_blocks=100, enable_caching=True)
@@ -100,7 +99,6 @@ def test_sliding_window_possible_cached_prefix():
         head_size=1,
         dtype=torch.float32,
         sliding_window=4,
-        use_mla=False,
     )
 
     block_pool = BlockPool(num_gpu_blocks=100, enable_caching=True)
@@ -165,7 +163,6 @@ def test_chunked_local_attention_remove_skipped_blocks():
         head_size=1,
         dtype=torch.float32,
         attention_chunk_size=4,
-        use_mla=False,
     )
 
     block_pool = BlockPool(num_gpu_blocks=2000, enable_caching=True)
@@ -217,7 +214,6 @@ def test_sliding_window_remove_skipped_blocks():
         head_size=1,
         dtype=torch.float32,
         sliding_window=4,
-        use_mla=False,
     )
 
     block_pool = BlockPool(num_gpu_blocks=2000, enable_caching=True)
@@ -285,7 +281,6 @@ def test_get_num_blocks_to_allocate():
         head_size=1,
         dtype=torch.float32,
         sliding_window=4,  # Placeholder value, not related to test result
-        use_mla=False,
     )
 
     block_pool = BlockPool(num_gpu_blocks=100, enable_caching=True)
@@ -308,7 +303,6 @@ def test_chunked_local_attention_get_num_blocks_to_allocate():
         head_size=1,
         dtype=torch.float32,
         attention_chunk_size=4,  # Placeholder value, not related to test result
-        use_mla=False,
     )
 
     block_pool = BlockPool(num_gpu_blocks=100, enable_caching=True)

tests/v1/engine/test_engine_core_client.py

@@ -836,8 +836,7 @@ def test_engine_core_proc_instantiation_cuda_empty(
         mock_spec = FullAttentionSpec(block_size=16,
                                       num_kv_heads=1,
                                       head_size=64,
-                                      dtype=torch.float16,
-                                      use_mla=False)
+                                      dtype=torch.float16)
 
         mock_executor.get_kv_cache_specs.return_value = [{
             "default": mock_spec

tests/v1/worker/test_gpu_model_runner.py

@@ -39,7 +39,6 @@ def initialize_kv_cache(runner: GPUModelRunner):
             runner.parallel_config),
         head_size=runner.model_config.get_head_size(),
         dtype=runner.kv_cache_dtype,
-        use_mla=False,
     )
     tensor_size = attn_spec.page_size_bytes * NUM_BLOCKS
     kv_cache_config = KVCacheConfig(

vllm/_custom_ops.py

@@ -1678,6 +1678,15 @@ def cp_gather_cache(src_cache: torch.Tensor,
                                            cu_seq_lens, batch_size, seq_starts)
 
 
+def indexer_k_quant_and_cache(k: torch.Tensor, kv_cache: torch.Tensor,
+                              slot_mapping: torch.Tensor,
+                              quant_block_size: int,
+                              kv_cache_dtype: str) -> None:
+    torch.ops._C_cache_ops.indexer_k_quant_and_cache(k, kv_cache, slot_mapping,
+                                                     quant_block_size,
+                                                     kv_cache_dtype)
+
+
 def get_device_attribute(attribute: int, device: int) -> int:
     return torch.ops._C_cuda_utils.get_device_attribute(attribute, device)
 

vllm/attention/backends/abstract.py

@@ -70,6 +70,7 @@ class AttentionBackend(ABC):
         block_size: int,
         num_kv_heads: int,
         head_size: int,
+        cache_dtype_str: str = "auto",
     ) -> Tuple[int, ...]:
         raise NotImplementedError
 

vllm/attention/layer.py

@@ -95,6 +95,7 @@ class Attention(nn.Module, AttentionLayerBase):
         logits_soft_cap: Optional[float] = None,
         per_layer_sliding_window: Optional[int] = None,
         use_mla: bool = False,
+        use_sparse: bool = False,
         prefix: str = "",
         attn_type: str = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[str] = None,
@@ -155,6 +156,7 @@ class Attention(nn.Module, AttentionLayerBase):
         self._o_scale_float: Optional[float] = None
 
         self.use_mla = use_mla
+        self.use_sparse = use_sparse
         self.num_heads = num_heads
         self.head_size = head_size
         self.num_kv_heads = num_kv_heads
@@ -187,7 +189,8 @@ class Attention(nn.Module, AttentionLayerBase):
                                                  kv_cache_dtype,
                                                  block_size,
                                                  use_mla=use_mla,
-                                                 has_sink=self.has_sink)
+                                                 has_sink=self.has_sink,
+                                                 use_sparse=use_sparse)
         else:
             self.attn_backend = attn_backend
 

vllm/attention/ops/common.py

@@ -138,3 +138,208 @@ def cp_lse_ag_out_rs(cp_attn_out: torch.Tensor,
     out, _ = correct_attn_out(cp_attn_out, lses, cp_group.rank_in_group, ctx)
     out = cp_group.reduce_scatter(out, dim=1)
     return out
+
+
+@triton.jit
+def _pack_seq_kernel(
+        x_ptr,  # [N, D]
+        out_ptr,  # [B, Lmax, D]
+        lengths_ptr,  # *i32, [B]
+        N: tl.constexpr,
+        D: tl.constexpr,
+        Lmax: tl.constexpr,
+        PAD_VALUE: tl.constexpr,
+        BLOCK_T: tl.constexpr,  # timesteps per program
+        BLOCK_D: tl.constexpr  # features per program
+):
+    pid_b = tl.program_id(0)  # batch id
+    pid_t = tl.program_id(1)  # block over time dimension
+    pid_d = tl.program_id(2)  # block over feature dimension
+    off_t = pid_t * BLOCK_T + tl.arange(0, BLOCK_T)  # [BLOCK_T]
+    off_d = pid_d * BLOCK_D + tl.arange(0, BLOCK_D)  # [BLOCK_D]
+
+    # Compute start index and sequence length from cumulative lengths
+    in_start = 0
+    for i in range(pid_b):
+        in_start += tl.load(lengths_ptr + i)
+    seq_len = tl.load(lengths_ptr + pid_b)
+
+    # valid time positions for this block
+    t_mask = off_t < Lmax
+
+    # compute input row indices for valid (b, t)
+    in_row = in_start + off_t
+    valid_row = (off_t < seq_len) & t_mask
+
+    # Pointers
+    # x_ptr: row-major [N, D]
+    x_row_ptr = x_ptr + in_row[:, None] * D + off_d[None, :]
+
+    # out_ptr: row-major [B, Lmax, D]
+    out_row_ptr = out_ptr + (pid_b * Lmax + off_t)[:,
+                                                   None] * D + off_d[None, :]
+
+    # Initialize with PAD (cast will occur as needed based on out_ptr dtype)
+    d_mask = off_d[None, :] < D
+    pad_vals = tl.full([BLOCK_T, BLOCK_D], PAD_VALUE, tl.float32)
+    tl.store(out_row_ptr, pad_vals, mask=t_mask[:, None] & d_mask)
+
+    # Load & write only where within seq_len
+    x_vals = tl.load(x_row_ptr, mask=valid_row[:, None] & d_mask)
+    tl.store(out_row_ptr, x_vals, mask=valid_row[:, None] & d_mask)
+
+
+def pack_seq_triton(x: torch.Tensor,
+                    lengths: torch.Tensor,
+                    pad_value: float = -float('inf'),
+                    block_t: int = 64,
+                    block_d: int = 64) -> torch.Tensor:
+    """
+    Pack sequences of different lengths into a batched tensor.
+    
+    Args:
+        x: [N, ...] - input tensor where N is total number of tokens
+        lengths: [B] - sequence lengths for each batch
+        pad_value: value to use for padding
+        block_t: block size for time dimension
+        block_d: block size for feature dimension
+        
+    Returns:
+        packed: [B, Lmax, ...] - packed tensor
+    """
+
+    # Handle multi-dimensional input by reshaping to (N, -1)
+    original_shape = x.shape
+    if len(original_shape) > 2:
+        N = original_shape[0]
+        x_reshaped = x.reshape(N, -1)
+        D = x_reshaped.shape[1]
+    else:
+        N, D = x.shape
+        x_reshaped = x
+
+    B = lengths.numel()
+    Lmax = int(lengths.max().item())
+
+    # Starts are computed inside the kernel from lengths
+
+    out = torch.empty((B, Lmax, D), device=x.device, dtype=x.dtype)
+
+    grid = (B, triton.cdiv(Lmax, block_t), triton.cdiv(D, block_d))
+    _pack_seq_kernel[grid](x_reshaped,
+                           out,
+                           lengths.int(),
+                           N,
+                           D,
+                           Lmax,
+                           PAD_VALUE=float(pad_value),
+                           BLOCK_T=block_t,
+                           BLOCK_D=block_d,
+                           num_warps=4,
+                           num_stages=2)
+
+    # Reshape output back to original dimensions (except first dimension)
+    if len(original_shape) > 2:
+        output_shape = (B, Lmax) + original_shape[1:]
+        out = out.reshape(output_shape)
+
+    return out
+
+
+@triton.jit
+def _unpack_seq_triton_kernel(
+        packed_ptr,  # [B, Lmax, D]
+        out_ptr,  # [N, D]
+        lengths_ptr,  # *i32, [B]
+        B: tl.constexpr,
+        Lmax: tl.constexpr,
+        D: tl.constexpr,
+        BLOCK_T: tl.constexpr,  # timesteps per program
+        BLOCK_D: tl.constexpr  # features per program
+):
+    pid_b = tl.program_id(0)  # batch id
+    pid_t = tl.program_id(1)  # block over time dimension
+    pid_d = tl.program_id(2)  # block over feature dimension
+    off_t = pid_t * BLOCK_T + tl.arange(0, BLOCK_T)  # [BLOCK_T]
+    off_d = pid_d * BLOCK_D + tl.arange(0, BLOCK_D)  # [BLOCK_D]
+
+    # bounds: compute start from cumulative lengths
+    in_start = 0
+    for i in range(pid_b):
+        in_start += tl.load(lengths_ptr + i)
+    seq_len = tl.load(lengths_ptr + pid_b)
+
+    # valid time positions for this block
+    t_mask = off_t < Lmax
+    valid_row = (off_t < seq_len) & t_mask
+
+    # compute output row indices for valid (b, t)
+    out_row = in_start + off_t
+
+    # Pointers
+    # packed_ptr: row-major [B, Lmax, D]
+    packed_row_ptr = packed_ptr + (pid_b * Lmax +
+                                   off_t)[:, None] * D + off_d[None, :]
+
+    # out_ptr: row-major [N, D]
+    out_row_ptr = out_ptr + out_row[:, None] * D + off_d[None, :]
+
+    # Load from packed tensor and store to output
+    d_mask = off_d[None, :] < D
+    packed_vals = tl.load(packed_row_ptr, mask=valid_row[:, None] & d_mask)
+    tl.store(out_row_ptr, packed_vals, mask=valid_row[:, None] & d_mask)
+
+
+def unpack_seq_triton(packed_tensor: torch.Tensor,
+                      lengths: torch.Tensor,
+                      block_t: int = 64,
+                      block_d: int = 64) -> torch.Tensor:
+    """
+    Unpack a packed decode query tensor back to the original format.
+    Efficient Triton implementation.
+    
+    Args:
+        packed_tensor: [B, Lmax, ...] - packed tensor from pack_seq_triton
+        lengths: [B] - sequence lengths for each batch
+        block_t: block size for time dimension
+        block_d: block size for feature dimension
+        
+    Returns:
+        unpacked_tensor: [N, ...] where N = sum(lengths)
+    """
+
+    # Handle multi-dimensional input by reshaping to (B, Lmax, -1)
+    original_shape = packed_tensor.shape
+    if len(original_shape) > 3:
+        B, Lmax = original_shape[:2]
+        packed_reshaped = packed_tensor.reshape(B, Lmax, -1)
+        D = packed_reshaped.shape[2]
+    else:
+        B, Lmax, D = packed_tensor.shape
+        packed_reshaped = packed_tensor
+
+    # Calculate total number of elements
+    N = int(lengths.sum().item())
+
+    out = torch.empty((N, D),
+                      device=packed_tensor.device,
+                      dtype=packed_tensor.dtype)
+
+    grid = (B, triton.cdiv(Lmax, block_t), triton.cdiv(D, block_d))
+    _unpack_seq_triton_kernel[grid](packed_reshaped,
+                                    out,
+                                    lengths.int(),
+                                    B,
+                                    Lmax,
+                                    D,
+                                    BLOCK_T=block_t,
+                                    BLOCK_D=block_d,
+                                    num_warps=4,
+                                    num_stages=2)
+
+    # Reshape output back to original dimensions (except first dimension)
+    if len(original_shape) > 3:
+        output_shape = (N, ) + original_shape[2:]
+        out = out.reshape(output_shape)
+
+    return out

vllm/attention/ops/flashmla.py

@@ -19,6 +19,15 @@ if current_platform.is_cuda():
 else:
     _flashmla_C_AVAILABLE = False
 
+if current_platform.is_cuda():
+    try:
+        import vllm._flashmla_extension_C  # noqa: F401
+        _flashmla_extension_C_AVAILABLE = True
+    except ImportError:
+        _flashmla_extension_C_AVAILABLE = False
+else:
+    _flashmla_extension_C_AVAILABLE = False
+
 
 def is_flashmla_supported() -> Tuple[bool, Optional[str]]:
     """
@@ -38,23 +47,33 @@ def is_flashmla_supported() -> Tuple[bool, Optional[str]]:
 
 def get_mla_metadata(
         cache_seqlens: torch.Tensor,
-    num_heads_per_head_k: int,
+        num_q_tokens_per_head_k: int,
         num_heads_k: int,
-) -> Tuple[torch.Tensor, torch.Tensor]:
+        num_heads_q: Optional[int] = None,
+        is_fp8_kvcache: bool = False,
+        topk: Optional[int] = None) -> Tuple[torch.Tensor, torch.Tensor]:
     """
     Arguments:
-        cache_seqlens: (batch_size), dtype torch.int32.
-        num_heads_per_head_k: Equals to seq_len_q * num_heads_q // num_heads_k.
-        num_heads_k: num_heads_k.
-
-    Return:
-        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), 
-                                 dtype torch.int32.
-        num_splits: (batch_size + 1), dtype torch.int32.
+    - cache_seqlens: (batch_size), dtype torch.int32.
+    - num_q_tokens_per_head_k: 
+            Equals to num_q_tokens_per_q_seq * num_heads_q // num_heads_k.
+    - num_heads_k: The number of k heads.
+    - num_heads_q: 
+            The number of q heads. 
+            This argument is optional when sparse attention is not enabled
+    - is_fp8_kvcache: Whether the k_cache and v_cache are in fp8 format.
+    - topk: If not None, sparse attention will be enabled, 
+            and only tokens in the `indices` array 
+            passed to `flash_mla_with_kvcache_sm90` will be attended to.
+
+    Returns:
+    - tile_scheduler_metadata: 
+            (num_sm_parts, TileSchedulerMetaDataSize), dtype torch.int32.
+    - num_splits: (batch_size + 1), dtype torch.int32.
     """
-    return torch.ops._flashmla_C.get_mla_metadata(cache_seqlens,
-                                                  num_heads_per_head_k,
-                                                  num_heads_k)
+    return torch.ops._flashmla_C.get_mla_decoding_metadata(
+        cache_seqlens, num_q_tokens_per_head_k, num_heads_k, num_heads_q,
+        is_fp8_kvcache, topk)
 
 
 def flash_mla_with_kvcache(
@@ -69,45 +88,95 @@ def flash_mla_with_kvcache(
     causal: bool = False,
     descale_q: Optional[torch.Tensor] = None,
     descale_k: Optional[torch.Tensor] = None,
+    is_fp8_kvcache: bool = False,
+    indices: Optional[torch.Tensor] = None,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     """
     Arguments:
-        q: (batch_size, seq_len_q, num_heads_q, head_dim).
-        k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
-        block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
-        cache_seqlens: (batch_size), torch.int32.
-        head_dim_v: Head_dim of v.
-        tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), 
-                                 torch.int32, return by get_mla_metadata.
-        num_splits: (batch_size + 1), torch.int32, return by get_mla_metadata.
-        softmax_scale: float. The scaling of QK^T before applying softmax. 
+    - q: (batch_size, seq_len_q, num_heads_q, head_dim).
+    - k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
+    - block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
+    - cache_seqlens: (batch_size), torch.int32.
+    - head_dim_v: Head dimension of v.
+    - tile_scheduler_metadata: 
+        (num_sm_parts, TileSchedulerMetaDataSize), torch.int32, 
+        returned by get_mla_metadata.
+    - num_splits: 
+        (batch_size + 1), torch.int32, returned by get_mla_metadata.
+    - softmax_scale: float. 
+        The scale of QK^T before applying softmax. 
         Default to 1 / sqrt(head_dim).
-        causal: bool. Whether to apply causal attention mask.
-        descale_q: (batch_size), torch.float32. Descaling factors for Q.
-        descale_k: (batch_size), torch.float32. Descaling factors for K.
+    - causal: bool. Whether to apply causal attention mask.
+    - descale_q: (batch_size), 
+        torch.float32. Descaling factors for Q, used for fp8 quantization.
+    - descale_k: (batch_size), 
+        torch.float32. Descaling factors for K, used for fp8 quantization.
+    - is_fp8_kvcache: bool. 
+        Whether the k_cache and v_cache are in fp8 format. 
+        For the format of FP8 KV cache, please refer to README.md
+    - indices: (batch_size, seq_len_q, topk), torch.int32. 
+        If not None, sparse attention will be enabled, 
+        and only tokens in the `indices` array will be attended to. 
+        Invalid indices should be set to -1 or numbers >= total_seq_len_kv. 
+        For details about how to set up `indices`, please refer to README.md.
 
-    Return:
-        out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
-        softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
+    Returns:
+    - out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
+    - softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
     """
     if softmax_scale is None:
         softmax_scale = q.shape[-1]**(-0.5)
+    if indices is not None:
+        # NOTE (zyongye): sparse attention is also causal
+        # since it only attend to the tokens before
+        # but here `causal` should not be specified
+        assert not causal, \
+            "causal must be `false` if sparse attention is enabled."
+    assert (descale_q is None) == (
+        descale_k is None
+    ), "descale_q and descale_k should be both None or both not None"
+
+    if (descale_q is not None) and (descale_k is not None):
+        out, softmax_lse = torch.ops._flashmla_extension_C.fwd_kvcache_mla_fp8(
+            q, k_cache, head_dim_v, cache_seqlens, block_table, softmax_scale,
+            causal, tile_scheduler_metadata, num_splits, descale_q, descale_k)
+    else:
         out, softmax_lse = torch.ops._flashmla_C.fwd_kvcache_mla(
-        q,
-        k_cache,
-        head_dim_v,
-        cache_seqlens,
-        block_table,
-        softmax_scale,
-        causal,
-        tile_scheduler_metadata,
-        num_splits,
-        descale_q,
-        descale_k,
-    )
-
-    # Note(hc): need revisit when we support DCP with decode query_len > 1.
-    return out.squeeze(1), softmax_lse.squeeze(-1)
+            q, k_cache, head_dim_v, cache_seqlens, block_table, softmax_scale,
+            causal, tile_scheduler_metadata, num_splits, is_fp8_kvcache,
+            indices)
+    return out, softmax_lse
+
+
+def flash_mla_sparse_prefill(
+    q: torch.Tensor,
+    kv: torch.Tensor,
+    indices: torch.Tensor,
+    sm_scale: float,
+    d_v: int = 512,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """
+    Sparse attention prefill kernel
+
+    Args:
+    - q: [s_q, h_q, d_qk], bfloat16
+    - kv: [s_kv, h_kv, d_qk], bfloat16
+    - indices: [s_q, h_kv, topk], int32. 
+        Invalid indices should be set to -1 or numbers >= s_kv
+    - sm_scale: float
+    - d_v: The dimension of value vectors. Can only be 512
+
+    Returns:
+    - (output, max_logits, lse)
+        About the definition of output, 
+        max_logits and lse, please refer to README.md
+    - output: [s_q, h_q, d_v], bfloat16
+    - max_logits:  [s_q, h_q], float
+    - lse: [s_q, h_q], float, 2-based log-sum-exp
+    """
+    results = torch.ops._flashmla_C.sparse_prefill_fwd(q, kv, indices,
+                                                       sm_scale, d_v)
+    return results
 
 
 #

vllm/attention/ops/paged_attn.py

@@ -50,6 +50,7 @@ class PagedAttention:
         block_size: int,
         num_kv_heads: int,
         head_size: int,
+        cache_dtype_str: str = "auto",
     ) -> Tuple[int, ...]:
         return (2, num_blocks, block_size * num_kv_heads * head_size)
 

vllm/attention/selector.py

@@ -144,6 +144,7 @@ def get_attn_backend(
     block_size: int,
     use_mla: bool = False,
     has_sink: bool = False,
+    use_sparse: bool = False,
 ) -> type[AttentionBackend]:
     """Selects which attention backend to use and lazily imports it."""
     # Accessing envs.* behind an @lru_cache decorator can cause the wrong
@@ -158,6 +159,7 @@ def get_attn_backend(
         use_v1=envs.VLLM_USE_V1,
         use_mla=use_mla,
         has_sink=has_sink,
+        use_sparse=use_sparse,
     )
 
 
@@ -170,6 +172,7 @@ def _cached_get_attn_backend(
     use_v1: bool = False,
     use_mla: bool = False,
     has_sink: bool = False,
+    use_sparse: bool = False,
 ) -> type[AttentionBackend]:
 
     # Check whether a particular choice of backend was
@@ -203,7 +206,7 @@ def _cached_get_attn_backend(
     # get device-specific attn_backend
     attention_cls = current_platform.get_attn_backend_cls(
         selected_backend, head_size, dtype, kv_cache_dtype, block_size, use_v1,
-        use_mla, has_sink)
+        use_mla, has_sink, use_sparse)
     if not attention_cls:
         raise ValueError(
             f"Invalid attention backend for {current_platform.device_name}")

vllm/config/cache.py

@@ -22,7 +22,8 @@ else:
 logger = init_logger(__name__)
 
 BlockSize = Literal[1, 8, 16, 32, 64, 128]
-CacheDType = Literal["auto", "fp8", "fp8_e4m3", "fp8_e5m2", "fp8_inc"]
+CacheDType = Literal["auto", "bfloat16", "fp8", "fp8_e4m3", "fp8_e5m2",
+                     "fp8_inc"]
 MambaDType = Literal["auto", "float32"]
 PrefixCachingHashAlgo = Literal["sha256", "sha256_cbor"]
 
@@ -52,7 +53,11 @@ class CacheConfig:
     cache_dtype: CacheDType = "auto"
     """Data type for kv cache storage. If "auto", will use model data type.
     CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. ROCm (AMD GPU) supports
-    fp8 (=fp8_e4m3). Intel Gaudi (HPU) supports fp8 (using fp8_inc)."""
+    fp8 (=fp8_e4m3). Intel Gaudi (HPU) supports fp8 (using fp8_inc).
+    Some models (namely DeepSeekV3.2) default to fp8, set to bfloat16 to use
+    bfloat16 instead, this is an invalid option for models that do not default
+    to fp8.
+    """
     is_attention_free: bool = False
     """Whether the model is attention-free. This is primarily set in
     `ModelConfig` and that value should be manually duplicated here."""
@@ -171,6 +176,7 @@ class CacheConfig:
         if self.cache_dtype == "auto":
             pass
         elif self.cache_dtype in get_args(CacheDType):
+            if self.cache_dtype.startswith("fp8"):
                 logger.info(
                     "Using fp8 data type to store kv cache. It reduces the GPU "
                     "memory footprint and boosts the performance. "

vllm/config/compilation.py

@@ -360,6 +360,7 @@ class CompilationConfig:
         "vllm.linear_attention",
         "vllm.plamo2_mamba_mixer",
         "vllm.gdn_attention",
+        "vllm.sparse_attn_indexer",
     ]
 
     def compute_hash(self) -> str:

vllm/config/model.py

@@ -1077,14 +1077,14 @@ class ModelConfig:
         if not hasattr(self.hf_text_config, "model_type"):
             return False
         elif self.hf_text_config.model_type in \
-            ('deepseek_v2', 'deepseek_v3', 'deepseek_mtp',
+            ('deepseek_v2', 'deepseek_v3', 'deepseek_v32', 'deepseek_mtp',
               'kimi_k2', 'longcat_flash'):
             return self.hf_text_config.kv_lora_rank is not None
         elif self.hf_text_config.model_type == 'eagle':
             # if the model is an EAGLE module, check for the
             # underlying architecture
             return self.hf_text_config.model.model_type in \
-                    ('deepseek_v2', 'deepseek_v3') \
+                    ('deepseek_v2', 'deepseek_v3', 'deepseek_v32') \
                 and self.hf_text_config.kv_lora_rank is not None
         return False
 

vllm/config/speculative.py

@@ -143,7 +143,7 @@ class SpeculativeConfig:
 
     @staticmethod
     def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
-        if hf_config.model_type == "deepseek_v3":
+        if hf_config.model_type in ("deepseek_v3", "deepseek_v32"):
             hf_config.model_type = "deepseek_mtp"
         if hf_config.model_type == "deepseek_mtp":
             n_predict = getattr(hf_config, "num_nextn_predict_layers", None)

vllm/model_executor/layers/layernorm.py

@@ -5,6 +5,7 @@ from typing import Optional, Union
 
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 
 import vllm.envs as envs
 from vllm.model_executor.custom_op import CustomOp
@@ -375,3 +376,20 @@ class PolyNorm(CustomOp):
         x: torch.Tensor,
     ) -> torch.Tensor:
         return poly_norm(x, self.weight, self.bias, self.variance_epsilon)
+
+
+class LayerNorm(nn.Module):
+    """
+    Layer Normalization.
+    """
+
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.dim = dim
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.bias = nn.Parameter(torch.zeros(dim, dtype=torch.float32))
+
+    def forward(self, x: torch.Tensor):
+        return F.layer_norm(x.float(), (self.dim, ), self.weight, self.bias,
+                            self.eps).type_as(x)

vllm/model_executor/layers/mla.py

@@ -24,6 +24,9 @@ class MLAModules:
     q_a_layernorm: Optional[torch.nn.Module]
     q_b_proj: Optional[torch.nn.Module]
     q_proj: Optional[torch.nn.Module]
+    indexer: Optional[torch.nn.Module]
+    is_sparse: bool
+    topk_indices_buffer: Optional[torch.Tensor]
 
 
 @CustomOp.register("multi_head_latent_attention")
@@ -76,6 +79,13 @@ class MultiHeadLatentAttention(CustomOp):
         self.kv_b_proj = mla_modules.kv_b_proj
         self.rotary_emb = mla_modules.rotary_emb
         self.o_proj = mla_modules.o_proj
+        self.indexer = mla_modules.indexer
+        self.is_sparse = mla_modules.is_sparse
+
+        if self.indexer is not None:
+            assert hasattr(self.indexer, "topk_tokens")
+            self.topk_tokens = self.indexer.topk_tokens
+            self.topk_indices_buffer = mla_modules.topk_indices_buffer
 
         # In the MLA backend, kv_cache includes both k_c and
         # pe (i.e. decoupled position embeddings). In particular,
@@ -92,6 +102,7 @@ class MultiHeadLatentAttention(CustomOp):
             quant_config=quant_config,
             prefix=f"{prefix}.attn",
             use_mla=True,
+            use_sparse=mla_modules.is_sparse,
             # MLA Args
             q_lora_rank=self.q_lora_rank,
             kv_lora_rank=self.kv_lora_rank,
@@ -100,6 +111,7 @@ class MultiHeadLatentAttention(CustomOp):
             qk_head_dim=self.qk_head_dim,
             v_head_dim=self.v_head_dim,
             kv_b_proj=self.kv_b_proj,
+            indexer=self.indexer,
         )
 
         self.prefix = prefix
@@ -145,6 +157,10 @@ class MultiHeadLatentAttention(CustomOp):
         q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
             positions, q[..., self.qk_nope_head_dim:], k_pe)
 
+        if self.indexer and self.is_sparse:
+            _topk_indices = self.indexer(hidden_states, q_c, positions,
+                                         self.rotary_emb)
+
         attn_out = self.mla_attn(
             q,
             kv_c_normed,

vllm/model_executor/models/config.py

@@ -346,8 +346,7 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
             block_size=1,
             num_kv_heads=model_config.get_num_kv_heads(parallel_config),
             head_size=model_config.get_head_size(),
-            dtype=kv_cache_dtype,
-            use_mla=model_config.use_mla).page_size_bytes
+            dtype=kv_cache_dtype).page_size_bytes
 
         model_cls, _ = ModelRegistry.resolve_model_cls(
             model_config.architecture,
@@ -401,6 +400,31 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
                 "exactly equal.", mamba_padding_pct)
 
 
+class DeepseekV3ForCausalLM(VerifyAndUpdateConfig):
+
+    @classmethod
+    def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
+        """
+        Updated fp8 cache to custom "fp8_ds_mla" format for DeepSeekV32
+        """
+        hf_config = vllm_config.model_config.hf_config
+
+        is_v32 = hasattr(hf_config, "index_topk")
+
+        if is_v32:
+            # For DeepSeekV3.2, we use a custom fp8 format as default (i.e.
+            #   "auto")
+            cache_config = vllm_config.cache_config
+            if cache_config.cache_dtype == "auto" or \
+                cache_config.cache_dtype.startswith("fp8"):
+                cache_config.cache_dtype = "fp8_ds_mla"
+                logger.info(
+                    "Using custom fp8 kv-cache format for DeepSeekV3.2")
+            if cache_config.cache_dtype == "bfloat16":
+                cache_config.cache_dtype = "auto"
+                logger.info("Using bfloat16 kv-cache for DeepSeekV3.2")
+
+
 MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
     "GteModel": SnowflakeGteNewModelConfig,
     "GteNewModel": GteNewModelConfig,
@@ -417,4 +441,5 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
     "MambaForCausalLM": MambaModelConfig,
     "Mamba2ForCausalLM": MambaModelConfig,
     "FalconMambaForCausalLM": MambaModelConfig,
+    "DeepseekV3ForCausalLM": DeepseekV3ForCausalLM,
 }

vllm/model_executor/models/deepseek_mtp.py

@@ -53,8 +53,20 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
         self.eh_proj = nn.Linear(config.hidden_size * 2,
                                  config.hidden_size,
                                  bias=False)
+
+        self.is_v32 = hasattr(config, "index_topk")
+        if self.is_v32:
+            topk_tokens = config.index_topk
+            topk_indices_buffer = torch.empty(
+                vllm_config.scheduler_config.max_num_batched_tokens,
+                topk_tokens,
+                dtype=torch.int32,
+                device="cuda")
+        else:
+            topk_indices_buffer = None
         self.shared_head = SharedHead(config=config, quant_config=quant_config)
-        self.mtp_block = DeepseekV2DecoderLayer(vllm_config, prefix)
+        self.mtp_block = DeepseekV2DecoderLayer(vllm_config, prefix,
+                                                topk_indices_buffer)
 
     def forward(
         self,

vllm/model_executor/models/longcat_flash.py

@@ -308,6 +308,7 @@ class FlashDecoderLayer(nn.Module):
 
     def __init__(
         self,
+        vllm_config: VllmConfig,
         config: FlashConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
@@ -329,6 +330,7 @@ class FlashDecoderLayer(nn.Module):
         # Dual attention structure
         self.self_attn = nn.ModuleList([
             DeepseekV2MLAAttention(
+                vllm_config=vllm_config,
                 config=config,
                 hidden_size=self.hidden_size,
                 num_heads=config.num_attention_heads,
@@ -454,6 +456,7 @@ class FlashModel(nn.Module):
         self.start_layer, self.end_layer, self.layers = make_layers(
             config.num_hidden_layers,
             lambda prefix: FlashDecoderLayer(
+                vllm_config,
                 config,
                 cache_config=cache_config,
                 quant_config=quant_config,