sglang 0.5.0rc2__py3-none-any.whl → 0.5.1__py3-none-any.whl

sglang/srt/layers/attention/utils.py

@@ -9,18 +9,89 @@ TRITON_PAD_NUM_PAGE_PER_BLOCK = 64
 
 @triton.jit
 def create_flashinfer_kv_indices_triton(
-    req_to_token_ptr,  # [max_batch, max_context_len]
+    req_to_token_ptr,
     req_pool_indices_ptr,
     page_kernel_lens_ptr,
     kv_indptr,
     kv_start_idx,
     kv_indices_ptr,
     req_to_token_ptr_stride: tl.constexpr,
+    PAGE_SIZE: tl.constexpr = 1,
 ):
+    """
+    Create KV indices for FlashInfer attention backend.
+
+    This Triton kernel builds a lookup table that maps from logical request/token
+    coordinates to physical token locations in the global KV cache pool. It's used
+    by FlashInfer attention backends to efficiently access scattered KV cache data.
+
+    The kernel processes each request in parallel and converts the req_to_token
+    lookup table into a flat list of token indices that can be used by attention kernels.
+
+    general idea:
+        blocktables/kv_indices_ptr = [batch_size * max_pages(for graph mode with
+                                                            fixed number of pages)]
+        max_pages = max_context_len / PAGED_SIZE
+        kv_indices_ptr will store the flat list of the pages used by each request
+    Args:
+        Inputs Arguments (non mutable):
+
+        req_to_token_ptr: Request to token location look up table
+                         Shape: [max_batch, max_context_len]
+        req_pool_indices_ptr: Request to pool index look up table. Each request uses
+                             one pool.
+                             Shape: [batch_size]
+        page_kernel_lens_ptr: sequence lengths per request
+                             Shape: [batch_size]
+        kv_indptr: Should be computed based on number of pages used by each request.
+                   It is used by flashinfer attention kernels to index into the kv_indices_ptr.
+                   per request.
+                  Shape: [batch_size + 1]
+                  kv_indptr[i] = start index in kv_indices for request i
+        kv_start_idx: Pointer to array containing start offsets for each request in SGL.
+                     Can be None. If provided, adds offset to token positions.
+
+        req_to_token_ptr_stride: Stride for the second dimension of req_to_token.
+                                Equal to max_context_len.
+
+        PAGED_SIZE: Number of tokens per page. Default is 1 for FlashInfer.
+
+        Outputs:
+        kv_indices_ptr: Pointer to output array where KV indices will be stored.
+                    Shape:[total-num-pages],
+                    where total_num_pages = sum(seq_lens // PAGED_SIZE)
+
+    Example:
+        If we have:
+        - req_pool_indices = [0, 1] (request 0 uses pool 0, request 1 uses pool 1)
+        - page_kernel_lens = [3, 2] (request 0 has 3 tokens, request 1 has 2 tokens)
+        - req_to_token = [[10, 11, 12, -1], [20, 21, -1, -1]] (tokens are the elements
+         in radix tree, use them as a pointer to the token location in the kv_indices_ptr)
+
+        The kernel will output:
+        If PAGE_SIZE = 1:
+        packed
+        - kv_indptr (passed in as input arg): [0,3,5]
+        - kv_indices = [10, 11, 12, 20, 21]
+        padded - max_pages is 10 tokens per req
+        - kv_indptr (passed in as input arg): [0,10, 20]
+        - kv_indices = [10, 11, 12, -1, -1, -1, -1, -1, -1, -1,
+                        20, 21, -1, -1, -1, -1, -1, -1, -1, -1]
+
+        If PAGE_SIZE = 2
+        packed:
+        - kv_indptr (passed in as input arg): [0,3,4]
+        - kv_indices = [5,6,10]
+        padded: max_pages is 4
+        - kv_indptr (passed in as input arg): [0,4,8,..] (note that 4 is the max_pages)
+        - kv_indices = [5, 6, -1, -1,
+                        10, -1, -1, -1]
+        This allows attention kernels to directly access the correct KV cache
+        entries for each request's tokens.
+    """
     BLOCK_SIZE: tl.constexpr = 512
+    NUM_PAGES_PER_BLOCK: tl.constexpr = BLOCK_SIZE // PAGE_SIZE
     pid = tl.program_id(axis=0)
-
-    # find the req pool idx, this is for batch to token
     req_pool_index = tl.load(req_pool_indices_ptr + pid)
     kv_indices_offset = tl.load(kv_indptr + pid)
 
@@ -31,19 +102,27 @@ def create_flashinfer_kv_indices_triton(
         kv_end = kv_start
     kv_end += tl.load(page_kernel_lens_ptr + pid).to(tl.int32)
 
-    num_loop = tl.cdiv(kv_end - kv_start, BLOCK_SIZE)
-    for i in range(num_loop):
-        # index into req_to_token_ptr needs to be int64
-        offset = tl.arange(0, BLOCK_SIZE).to(tl.int64) + i * BLOCK_SIZE
-        mask = offset < kv_end - kv_start
-        data = tl.load(
-            req_to_token_ptr
-            + req_pool_index * req_to_token_ptr_stride
-            + kv_start
-            + offset,
-            mask=mask,
+    kv_range = kv_end - kv_start
+    num_pages = tl.cdiv(kv_range, PAGE_SIZE)
+    num_loops = tl.cdiv(kv_range, BLOCK_SIZE)
+    req_to_token_block_start = (
+        req_to_token_ptr + req_pool_index * req_to_token_ptr_stride + kv_start
+    )
+    for i in range(num_loops):
+        token_offsets_in_block = (
+            tl.arange(0, NUM_PAGES_PER_BLOCK).to(tl.int64) + i * NUM_PAGES_PER_BLOCK
+        ) * PAGE_SIZE
+        page_offsets_in_block = token_offsets_in_block // PAGE_SIZE
+        valid_tokens = token_offsets_in_block < kv_range
+        valid_pages = page_offsets_in_block < num_pages
+        token_numbers = tl.load(
+            req_to_token_block_start + token_offsets_in_block, mask=valid_tokens
+        )
+        tl.store(
+            kv_indices_ptr + kv_indices_offset + page_offsets_in_block,
+            token_numbers // PAGE_SIZE,  # write the page numbers to kv_indices_ptr
+            mask=valid_pages,
         )
-        tl.store(kv_indices_ptr + kv_indices_offset + offset, data, mask=mask)
 
 
 @triton.jit

sglang/srt/layers/attention/vision.py

@@ -12,7 +12,12 @@ import torch.nn.functional as F
 from einops import rearrange
 
 from sglang.srt.layers.dp_attention import get_attention_tp_rank, get_attention_tp_size
-from sglang.srt.utils import is_cuda, print_info_once
+from sglang.srt.utils import (
+    get_device_capability,
+    is_blackwell,
+    is_cuda,
+    print_info_once,
+)
 
 _is_cuda = is_cuda()
 
@@ -20,7 +25,6 @@ if _is_cuda:
     from sgl_kernel.flash_attn import flash_attn_varlen_func
 
 from sglang.srt.distributed import (
-    parallel_state,
     split_tensor_along_last_dim,
     tensor_model_parallel_all_gather,
 )
@@ -402,18 +406,14 @@ class VisionAttention(nn.Module):
                 self.dummy_dim, eps=layer_norm_eps, var_hidden_size=embed_dim
             )
 
-        # priority: server_args > passed qkv_backend > sdpa
-        if global_server_args_dict["mm_attention_backend"] is None:
-            if qkv_backend is None:
-                if is_cuda():
-                    # Double prefill throughput by setting attn backend to Triton on CUDA
-                    qkv_backend = "triton_attn"
-                else:
-                    qkv_backend = "sdpa"
+        # Select attention backend via a unified method
+        _passed_backend = qkv_backend
+        qkv_backend = self._determine_attention_backend(_passed_backend)
+        if (
+            global_server_args_dict["mm_attention_backend"] is None
+            and _passed_backend is None
+        ):
             print_info_once(f"Multimodal attention backend not set. Use {qkv_backend}.")
-        else:
-            qkv_backend = global_server_args_dict["mm_attention_backend"]
-
         print_info_once(f"Using {qkv_backend} as multimodal attention backend.")
 
         self.customized_position_embedding_applier = (
@@ -461,6 +461,33 @@ class VisionAttention(nn.Module):
             prefix=add_prefix("proj", prefix),
         )
 
+    def _determine_attention_backend(self, passed_backend: Optional[str]) -> str:
+        """Decide the multimodal attention backend string.
+
+        Priority: server args override > constructor arg > platform default.
+
+        Platform defaults:
+        - CUDA: "triton_attn"
+        - Non-CUDA: "sdpa"
+        """
+        override_backend = global_server_args_dict["mm_attention_backend"]
+        if override_backend is not None:
+            backend = override_backend
+        elif passed_backend is not None:
+            backend = passed_backend
+        elif is_cuda():
+            major, minor = get_device_capability()
+            if major == 9:
+                backend = "fa3"
+            else:
+                backend = "triton_attn"
+        else:
+            backend = "sdpa"
+        if backend == "fa3" and is_blackwell():
+            raise ValueError("The 'fa3' backend is not supported on Blackwell GPUs")
+
+        return backend
+
     def _apply_qk_norm(self, q: torch.Tensor, k: torch.Tensor):
         """apply qk norm for internvl vit attn"""
         q = q.flatten(1, 2)

sglang/srt/layers/attention/vision_utils.py

@@ -0,0 +1,65 @@
+"""Utility functions for vision attention layers."""
+
+import torch
+
+from sglang.srt.layers.dp_attention import get_attention_tp_size
+
+
+def update_vit_attn_dummy_heads_config(config):
+    """Update HF config to ensure vision attention num_attention_heads is divisible by tp_size"""
+    tp_size = get_attention_tp_size()
+    num_heads = getattr(
+        config.vision_config,
+        "num_heads",
+        getattr(config.vision_config, "num_attention_heads", None),
+    )
+    head_dim = config.vision_config.hidden_size // num_heads
+    num_dummy_heads = 0
+
+    if num_heads % tp_size != 0:
+        num_dummy_heads = ((num_heads + tp_size - 1) // tp_size) * tp_size - num_heads
+
+    setattr(config.vision_config, "head_dim", head_dim)
+    setattr(config.vision_config, "num_dummy_heads", num_dummy_heads)
+
+
+def pad_vit_attn_dummy_heads(config, name: str, loaded_weight: torch.Tensor):
+    """Pad attention qkv weights for dummy heads"""
+    num_dummy_heads = config.vision_config.num_dummy_heads
+    if num_dummy_heads == 0:
+        return loaded_weight
+    head_dim = config.vision_config.head_dim
+
+    if "attn.qkv_proj" in name:
+        wq, wk, wv = loaded_weight.chunk(3, dim=0)
+        if name.endswith(".weight"):
+            dummy_shape = [num_dummy_heads, head_dim, wq.shape[-1]]
+        elif name.endswith(".bias"):
+            dummy_shape = [num_dummy_heads, head_dim]
+        else:
+            raise RuntimeError(f"Unsupported weight with name={name}")
+        pad_func = lambda x: torch.cat(
+            [x.unflatten(0, (-1, head_dim)), x.new_zeros(dummy_shape)], dim=0
+        ).flatten(0, 1)
+        wq, wk, wv = pad_func(wq), pad_func(wk), pad_func(wv)
+        loaded_weight = torch.cat([wq, wk, wv], dim=0)
+    elif any([_ in name for _ in ["attn.q_proj", "attn.k_proj", "attn.v_proj"]]):
+        if name.endswith(".weight"):
+            dummy_shape = [num_dummy_heads, head_dim, loaded_weight.shape[-1]]
+        elif name.endswith(".bias"):
+            dummy_shape = [num_dummy_heads, head_dim]
+        else:
+            raise RuntimeError(f"Unsupported weight with name={name}")
+        padded_weight = loaded_weight.new_zeros(dummy_shape)
+        loaded_weight = torch.cat(
+            [loaded_weight.unflatten(0, (-1, head_dim)), padded_weight], dim=0
+        ).flatten(0, 1)
+    elif "attn.proj.weight" in name:
+        padded_weight = loaded_weight.new_zeros(
+            loaded_weight.shape[0], head_dim * num_dummy_heads
+        )
+        loaded_weight = torch.cat([loaded_weight, padded_weight], dim=-1)
+    elif "attn.q_norm.weight" in name or "attn.k_norm.weight" in name:
+        padded_weight = loaded_weight.new_zeros(head_dim * num_dummy_heads)
+        loaded_weight = torch.cat([loaded_weight, padded_weight], dim=0)
+    return loaded_weight

sglang/srt/layers/communicator.py

@@ -17,7 +17,7 @@ from enum import Enum, auto
 from functools import partial
 from typing import Dict, Optional
 
-import torch.distributed
+import torch
 
 from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
@@ -34,6 +34,11 @@ from sglang.srt.layers.dp_attention import (
     get_attention_tp_size,
     get_global_dp_buffer,
     get_local_dp_buffer,
+    is_dp_attention_enabled,
+)
+from sglang.srt.layers.moe import (
+    get_moe_a2a_backend,
+    should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.utils import is_sm100_supported
 from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -43,6 +48,8 @@ from sglang.srt.utils import is_cuda, is_flashinfer_available
 _is_flashinfer_available = is_flashinfer_available()
 _is_sm100_supported = is_cuda() and is_sm100_supported()
 
+FUSE_ALLREDUCE_MAX_BATCH_SIZE = 2048
+
 
 class ScatterMode(Enum):
     """
@@ -111,7 +118,11 @@ class LayerScatterModes:
         if context.is_layer_sparse:
             return (
                 ScatterMode.SCATTERED
-                if not global_server_args_dict["moe_a2a_backend"].is_standard()
+                if (
+                    # Token dispatch/combine will be handled outside of LayerCommunicator for these modes.
+                    not get_moe_a2a_backend().is_none()
+                    or should_use_flashinfer_cutlass_moe_fp4_allgather()
+                )
                 else ScatterMode.FULL
             )
         else:
@@ -154,11 +165,13 @@ class LayerCommunicator:
         post_attention_layernorm: torch.nn.Module,
         # Reduce scatter requires skipping all-reduce in model code after MoE/MLP, so only enable for models which have that implemented. Remove flag once done for all models that use LayerCommunicator.
         allow_reduce_scatter: bool = False,
+        is_last_layer: bool = False,
     ):
         self.layer_scatter_modes = layer_scatter_modes
         self.input_layernorm = input_layernorm
         self.post_attention_layernorm = post_attention_layernorm
         self.allow_reduce_scatter = allow_reduce_scatter
+        self.is_last_layer = is_last_layer
 
         self._context = CommunicateContext.init_new()
         self._communicate_simple_fn = CommunicateSimpleFn.get_fn(
@@ -256,6 +269,41 @@ class LayerCommunicator:
             and forward_batch.dp_padding_mode.is_max_len()
         )
 
+    def should_fuse_mlp_allreduce_with_next_layer(
+        self, forward_batch: ForwardBatch
+    ) -> bool:
+        speculative_algo = global_server_args_dict.get("speculative_algorithm", None)
+        if (
+            is_dp_attention_enabled()
+            and speculative_algo is not None
+            and speculative_algo.is_eagle()
+        ):
+            return False
+
+        batch_size = (
+            forward_batch.input_ids.shape[0]
+            if hasattr(forward_batch, "input_ids")
+            else 0
+        )
+        if batch_size > FUSE_ALLREDUCE_MAX_BATCH_SIZE:
+            return False
+
+        static_conditions_met = (
+            (not self.is_last_layer)
+            and (self._context.tp_size > 1)
+            and global_server_args_dict.get("enable_flashinfer_allreduce_fusion", False)
+            and _is_flashinfer_available
+        )
+
+        if not static_conditions_met:
+            return False
+
+        return (
+            batch_size > 0
+            and batch_size <= FUSE_ALLREDUCE_MAX_BATCH_SIZE
+            and (not self.is_last_layer)
+        )
+
 
 @dataclass
 class CommunicateContext:
@@ -382,7 +430,7 @@ class CommunicateWithAllReduceAndLayerNormFn:
             )
 
         raise NotImplementedError(
-            f"{hidden_states_input_mode=} {residual_input_mode=} {residual_output_mode=} {residual_output_mode=}"
+            f"{hidden_states_input_mode=} {residual_input_mode=} {hidden_states_output_mode=} {residual_output_mode=}"
         )
 
     @staticmethod

sglang/srt/layers/dp_attention.py

@@ -72,6 +72,7 @@ class _DpGatheredBufferWrapper:
     _device: torch.device
     _global_dp_buffer_len: int
     _local_dp_buffer_len: int
+    _global_num_tokens: Optional[List[int]]
 
     @classmethod
     def set_metadata(cls, hidden_size: int, dtype: torch.dtype, device: torch.device):
@@ -80,9 +81,15 @@ class _DpGatheredBufferWrapper:
         cls._device = device
 
     @classmethod
-    def set_dp_buffer_len(cls, global_dp_buffer_len: int, local_dp_buffer_len: int):
+    def set_dp_buffer_len(
+        cls,
+        global_dp_buffer_len: int,
+        local_dp_buffer_len: int,
+        global_num_tokens: Optional[List[int]] = None,
+    ):
         cls._global_dp_buffer_len = global_dp_buffer_len
         cls._local_dp_buffer_len = local_dp_buffer_len
+        cls._global_num_tokens = global_num_tokens
 
     @classmethod
     def get_global_dp_buffer(cls) -> torch.Tensor:
@@ -108,10 +115,18 @@ class _DpGatheredBufferWrapper:
     def get_local_dp_buffer_len(cls) -> int:
         return cls._local_dp_buffer_len
 
+    @classmethod
+    def get_dp_global_num_tokens(cls) -> List[int]:
+        return cls._global_num_tokens
+
 
-def set_dp_buffer_len(global_dp_buffer_len: int, local_dp_buffer_len: int):
+def set_dp_buffer_len(
+    global_dp_buffer_len: int,
+    local_dp_buffer_len: int,
+    global_num_tokens: Optional[List[int]] = None,
+):
     _DpGatheredBufferWrapper.set_dp_buffer_len(
-        global_dp_buffer_len, local_dp_buffer_len
+        global_dp_buffer_len, local_dp_buffer_len, global_num_tokens
     )
 
 
@@ -131,6 +146,10 @@ def get_local_dp_buffer_len() -> int:
     return _DpGatheredBufferWrapper.get_local_dp_buffer_len()
 
 
+def get_dp_global_num_tokens() -> List[int]:
+    return _DpGatheredBufferWrapper.get_dp_global_num_tokens()
+
+
 def compute_dp_attention_world_info(enable_dp_attention, tp_rank, tp_size, dp_size):
     if not enable_dp_attention:
         return tp_rank, tp_size, 0
@@ -215,7 +234,7 @@ def initialize_dp_attention(
     _DpGatheredBufferWrapper.set_metadata(
         hidden_size=model_config.hidden_size,
         dtype=model_config.dtype,
-        device=torch.device("cuda"),
+        device=torch.device(server_args.device),
     )
 
 

sglang/srt/layers/elementwise.py

@@ -486,3 +486,97 @@ def gelu_and_mul_triton(
         return out_hidden_states, out_scales
     else:
         return out_hidden_states, None
+
+
+# silu on first half of vector
+@triton.jit
+def silu_and_mul_kernel(
+    out_hidden_states_ptr,  # (bs, hidden_dim)
+    out_scales_ptr,  # (bs,)
+    hidden_states_ptr,  # (bs, hidden_dim * 2)
+    quant_max: tl.constexpr,
+    static_scale: tl.constexpr,
+    hidden_dim: tl.constexpr,  # the output hidden_dim
+    BLOCK_SIZE: tl.constexpr,
+):
+    pid = tl.program_id(axis=0)
+
+    input_start = pid * hidden_dim * 2
+    output_start = pid * hidden_dim
+
+    input1_offs = tl.arange(0, BLOCK_SIZE)
+    mask = tl.arange(0, BLOCK_SIZE) < hidden_dim  # shared for input1, input3, output
+    input3_offs = hidden_dim + tl.arange(0, BLOCK_SIZE)
+    output_offs = tl.arange(0, BLOCK_SIZE)
+
+    x1 = tl.load(
+        hidden_states_ptr + input_start + input1_offs, mask=mask, other=0.0
+    ).to(tl.float32)
+    x3 = tl.load(
+        hidden_states_ptr + input_start + input3_offs, mask=mask, other=0.0
+    ).to(tl.float32)
+
+    # silu
+    # cast down before mul to better match training?
+    silu_x1 = x1 * tl.sigmoid(x1)
+    out = x3 * silu_x1.to(hidden_states_ptr.dtype.element_ty)
+
+    if quant_max is not None:
+        raise NotImplementedError()
+
+    tl.store(out_hidden_states_ptr + output_start + output_offs, out, mask=mask)
+
+
+def silu_and_mul_triton(
+    hidden_states,
+    scales=None,
+    quantize=None,  # dtype to quantize to
+    out=None,
+):
+    bs, in_hidden_dim = hidden_states.shape
+    hidden_dim = in_hidden_dim // 2
+
+    if out is None:
+        out_hidden_states = torch.empty(
+            (bs, hidden_dim),
+            dtype=quantize or hidden_states.dtype,
+            device=hidden_states.device,
+        )
+    else:
+        assert out.shape == (bs, hidden_dim)
+        assert out.dtype == (quantize or hidden_states.dtype)
+        out_hidden_states = out
+    out_scales = None
+    static_scale = False
+    if quantize is not None:
+        if scales is None:
+            out_scales = torch.empty(
+                (bs,), dtype=torch.float32, device=hidden_states.device
+            )
+        else:
+            out_scales = scales
+            static_scale = True
+
+    max_warps = 16 if _is_hip else 32
+    config = {
+        # 8 ele per thread (not tuned)
+        "num_warps": max(
+            min(triton.next_power_of_2(triton.cdiv(hidden_dim, 8 * 32)), max_warps), 4
+        ),
+    }
+
+    silu_and_mul_kernel[(bs,)](
+        out_hidden_states,
+        out_scales,
+        hidden_states,
+        quant_max=torch.finfo(quantize).max if quantize is not None else None,
+        static_scale=static_scale,
+        hidden_dim=hidden_dim,
+        BLOCK_SIZE=triton.next_power_of_2(hidden_dim),
+        **config,
+    )
+
+    if quantize is not None:
+        return out_hidden_states, out_scales
+    else:
+        return out_hidden_states, None

sglang/srt/layers/flashinfer_comm_fusion.py

@@ -5,7 +5,11 @@ import torch
 import torch.distributed as dist
 
 from sglang.srt.distributed import get_tensor_model_parallel_world_size
-from sglang.srt.utils import is_flashinfer_available
+from sglang.srt.utils import (
+    direct_register_custom_op,
+    is_flashinfer_available,
+    supports_custom_op,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -196,6 +200,30 @@ def flashinfer_allreduce_residual_rmsnorm(
     return norm_out, residual_out
 
 
+def fake_flashinfer_allreduce_residual_rmsnorm(
+    input_tensor: torch.Tensor,
+    residual: torch.Tensor,
+    weight: torch.Tensor,
+    eps: float = 1e-6,
+    max_token_num: int = 2048,
+    use_oneshot: Optional[bool] = None,
+    trigger_completion_at_end: bool = False,
+    fp32_acc: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    residual_out = torch.empty_like(residual)
+    norm_out = torch.empty_like(input_tensor)
+    return norm_out, residual_out
+
+
+if supports_custom_op():
+    direct_register_custom_op(
+        "flashinfer_allreduce_residual_rmsnorm",
+        flashinfer_allreduce_residual_rmsnorm,
+        mutates_args=["input_tensor", "residual", "weight"],
+        fake_impl=fake_flashinfer_allreduce_residual_rmsnorm,
+    )
+
+
 def cleanup_flashinfer_workspace():
     global _workspace_manager
     if _workspace_manager is not None:

sglang/srt/layers/layernorm.py

@@ -27,6 +27,7 @@ from sglang.srt.utils import (
     is_cuda,
     is_hip,
     is_npu,
+    supports_custom_op,
 )
 
 _is_cuda = is_cuda()
@@ -202,8 +203,14 @@ class RMSNorm(CustomOp):
                 flashinfer_allreduce_residual_rmsnorm,
             )
 
+            fused_op = (
+                torch.ops.sglang.flashinfer_allreduce_residual_rmsnorm
+                if supports_custom_op()
+                else flashinfer_allreduce_residual_rmsnorm
+            )
+
             if get_tensor_model_parallel_world_size() > 1:
-                fused_result = flashinfer_allreduce_residual_rmsnorm(
+                fused_result = fused_op(
                     input_tensor=x,
                     residual=residual,
                     weight=self.weight,

sglang/srt/layers/linear.py

@@ -110,6 +110,20 @@ def adjust_scalar_to_fused_array(param, loaded_weight, shard_id):
     return param[shard_id], loaded_weight
 
 
+def adjust_shard_offsets(shard_offsets, loaded_weight, dim):
+    actual_weight_size = loaded_weight.size(dim)
+    target_weight_size = shard_offsets[-1][-1] + shard_offsets[-1][-2]
+    if actual_weight_size != target_weight_size:
+        new_shard_offsets = []
+        new_offset = 0
+        for shard_id, shard_offset, shard_size in shard_offsets:
+            actual_shard_size = actual_weight_size * shard_size // target_weight_size
+            new_shard_offsets.append((shard_id, new_offset, actual_shard_size))
+            new_offset += actual_shard_size
+        return new_shard_offsets
+    return shard_offsets
+
+
 class LinearBase(torch.nn.Module):
     """Base linear layer.
 
@@ -535,6 +549,11 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
             packed_dim = getattr(param, "packed_dim", None)
 
             use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
+            if _is_cpu:
+                shard_offsets = adjust_shard_offsets(
+                    shard_offsets, loaded_weight, output_dim
+                )
+
             for shard_id, shard_offset, shard_size in shard_offsets:
                 # Special case for Quantization.
                 # If quantized, we need to adjust the offset and size to account
@@ -977,6 +996,11 @@ class QKVParallelLinear(ColumnParallelLinear):
             use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
 
             packed_dim = getattr(param, "packed_dim", None)
+            if _is_cpu:
+                shard_offsets = adjust_shard_offsets(
+                    shard_offsets, loaded_weight, output_dim
+                )
+
             for shard_id, shard_offset, shard_size in shard_offsets:
                 # Special case for Quantized Weights.
                 # If quantized, we need to adjust the offset and size to account

sglang/srt/layers/logits_processor.py

@@ -191,7 +191,11 @@ class LogitsMetadata:
         else:
             self.global_dp_buffer_len = self.global_dp_buffer_len
 
-        set_dp_buffer_len(self.global_dp_buffer_len, self.dp_local_num_tokens)
+        set_dp_buffer_len(
+            self.global_dp_buffer_len,
+            self.dp_local_num_tokens,
+            self.global_num_tokens_for_logprob_cpu,
+        )
 
 
 class LogitsProcessor(nn.Module):