sglang 0.4.9.post3__py3-none-any.whl → 0.4.9.post4__py3-none-any.whl

sglang/srt/function_call/kimik2_detector.py

@@ -18,16 +18,21 @@ logger = logging.getLogger(__name__)
 
 
 class KimiK2Detector(BaseFormatDetector):
+    """
+    Detector for Kimi K2 model function call format.
+
+    Format Structure:
+    ```
+    <|tool_calls_section_begin|>
+    <|tool_call_begin|>functions.{func_name}:{index} <|tool_call_argument_begin|>{json_args}<|tool_call_end|>
+    <|tool_calls_section_end|>
+    ```
+
+    Reference: https://huggingface.co/moonshotai/Kimi-K2-Instruct/blob/main/docs/tool_call_guidance.md
+    """
 
     def __init__(self):
         super().__init__()
-        self._buffer = ""
-        self.current_tool_name_sent: bool = False
-        self.prev_tool_call_arr: list[dict] = []
-        self.current_tool_id: int = -1
-        self.streamed_args_for_tool: list[str] = (
-            []
-        )  # map what has been streamed for each tool so far to a list
 
         self.bot_token: str = "<|tool_calls_section_begin|>"
         self.eot_token: str = "<|tool_calls_section_end|>"
@@ -114,11 +119,7 @@ class KimiK2Detector(BaseFormatDetector):
             return StreamingParseResult(normal_text=new_text)
 
         if not hasattr(self, "_tool_indices"):
-            self._tool_indices = {
-                tool.function.name: i
-                for i, tool in enumerate(tools)
-                if tool.function and tool.function.name
-            }
+            self._tool_indices = self._get_tool_indices(tools)
 
         calls: list[ToolCallItem] = []
         try:
@@ -150,7 +151,7 @@ class KimiK2Detector(BaseFormatDetector):
                         )
                     )
                     self.current_tool_name_sent = True
-                    # Store the tool call info for adapter.py
+                    # Store the tool call info for serving layer completions endpoint
                     self.prev_tool_call_arr[self.current_tool_id] = {
                         "name": function_name,
                         "arguments": {},
@@ -214,7 +215,31 @@ class KimiK2Detector(BaseFormatDetector):
             return StreamingParseResult(normal_text=current_text)
 
     def structure_info(self) -> _GetInfoFunc:
-        raise NotImplementedError()
+        """Return function that creates StructureInfo for guided generation."""
+
+        def get_info(name: str) -> StructureInfo:
+            return StructureInfo(
+                begin=f"<|tool_calls_section_begin|><|tool_call_begin|>functions.{name}:0 <|tool_call_argument_begin|>",
+                end="<|tool_call_end|><|tool_calls_section_end|>",
+                trigger="<|tool_calls_section_begin|>",
+            )
+
+        return get_info
 
-    def build_ebnf(self, tools: List[Tool]):
-        raise NotImplementedError()
+    def build_ebnf(self, tools: List[Tool]) -> str:
+        """
+        Build EBNF grammar for KimiK2 tool call format.
+
+        NOTE: The call_rule_fmt uses [0-9]+ for the function index to allow the grammar
+        to accept any numeric index (0, 1, 2, etc.) for proper sequential indexing in
+        multiple function call scenarios, while still maintaining the correct KimiK2
+        format structure for constrained generation.
+        """
+        return EBNFComposer.build_ebnf(
+            tools,
+            sequence_start_token=self.bot_token,
+            sequence_end_token=self.eot_token,
+            tool_call_separator="",
+            call_rule_fmt='"<|tool_call_begin|>functions.{name}:" [0-9]+ " <|tool_call_argument_begin|>" {arguments_rule} "<|tool_call_end|>"',
+            function_format="json",
+        )

sglang/srt/function_call/llama32_detector.py

@@ -16,9 +16,12 @@ logger = logging.getLogger(__name__)
 
 class Llama32Detector(BaseFormatDetector):
     """
-    Detector for Llama 3.2 models.
-    Assumes function call format:
-      <|python_tag|>{"name":"xxx", "arguments":{...}}
+    Detector for Llama 3.2 models with json tool call format.
+
+    Format Structure:
+    ```
+    <python_tag>{"name":"xxx", "arguments":{...}}
+    ```
     """
 
     def __init__(self):

sglang/srt/function_call/mistral_detector.py

@@ -17,9 +17,17 @@ logger = logging.getLogger(__name__)
 
 class MistralDetector(BaseFormatDetector):
     """
-    Detector for Mistral models.
-    Assumes function call format:
-      [TOOL_CALLS] [{"name":"func1", "arguments":{...}}, {"name":"func2", "arguments":{...}}]
+    Detector for Mistral model function call format.
+
+    The Mistral format uses a simple bracket-delimited structure with JSON arrays
+    containing function call objects.
+
+    Format Structure:
+    ```
+    [TOOL_CALLS] [{"name": "function_name", "arguments": {json_args}}, ...]
+    ```
+
+    Reference: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3?chat_template=default
     """
 
     def __init__(self):

sglang/srt/function_call/pythonic_detector.py

@@ -8,7 +8,6 @@ from sglang.srt.entrypoints.openai.protocol import Tool
 from sglang.srt.function_call.base_format_detector import BaseFormatDetector
 from sglang.srt.function_call.core_types import (
     StreamingParseResult,
-    StructureInfo,
     ToolCallItem,
     _GetInfoFunc,
 )
@@ -19,10 +18,17 @@ logger = logging.getLogger(__name__)
 
 class PythonicDetector(BaseFormatDetector):
     """
-    Detector for Llama-3.2 and Llama-4 models with pythonic tool call format.
-    Assumes function call format:
-      [tool1(arg1=val1, arg2=val2), tool2(arg1=val3)]
-    Arguments are Python literals (not JSON).
+    Detector for Llama-4 models with Pythonic tool call format.
+
+    The Pythonic format uses Python function call syntax within square brackets,
+    with arguments as Python literals rather than JSON.
+
+    Format Structure:
+    ```
+    [tool1(arg1=val1, arg2=val2), tool2(arg1=val3)]
+    ```
+
+    Reference: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct?chat_template=default
     """
 
     def __init__(self):
@@ -75,11 +81,7 @@ class PythonicDetector(BaseFormatDetector):
                 return StreamingParseResult(normal_text=normal_text, calls=[])
 
             calls = []
-            tool_indices = {
-                tool.function.name: i
-                for i, tool in enumerate(tools)
-                if tool.function.name
-            }
+            tool_indices = self._get_tool_indices(tools)
             for call_index, call in enumerate(parsed.elts):
                 if not isinstance(call.func, ast.Name):
                     continue
@@ -213,11 +215,11 @@ class PythonicDetector(BaseFormatDetector):
         else:
             raise ValueError("Tool call arguments must be literals")
 
-    def structure_info(self) -> _GetInfoFunc:
-        def info(name: str):
-            return StructureInfo(begin=f"[{name}(", end=")]", trigger=f"[{name}(")
+    def supports_structural_tag(self) -> bool:
+        return False
 
-        return info
+    def structure_info(self) -> _GetInfoFunc:
+        raise NotImplementedError
 
     def build_ebnf(self, tools: List[Tool]) -> Optional[str]:
         return EBNFComposer.build_ebnf(

sglang/srt/function_call/qwen25_detector.py

@@ -17,9 +17,18 @@ logger = logging.getLogger(__name__)
 
 class Qwen25Detector(BaseFormatDetector):
     """
-    Detector for Qwen 2.5 models.
-    Assumes function call format:
-      <tool_call>\n{"name":"func1", "arguments":{...}}\n</tool_call>\n<tool_call>\n{"name":"func2", "arguments":{...}}\n</tool_call>
+    Detector for Qwen 2.5 and Qwen 3 model function call format.
+
+    Format Structure:
+    ```
+    <tool_call>\n{"name":"func1", "arguments":{...}}\n</tool_call>\n<tool_call>\n{"name":"func2", "arguments":{...}}\n</tool_call>
+    ```
+
+    Key Components:
+    - Tool Call Tags: `<tool_call>` and `</tool_call>` wrap each individual call
+    - Function Call Object: JSON object with "name" and "arguments" fields
+
+    Reference: https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct?chat_template=default
     """
 
     def __init__(self):

sglang/srt/function_call/{qwen3_detector.py → qwen3_coder_detector.py}

@@ -9,7 +9,6 @@ from sglang.srt.entrypoints.openai.protocol import Tool
 from sglang.srt.function_call.base_format_detector import BaseFormatDetector
 from sglang.srt.function_call.core_types import (
     StreamingParseResult,
-    StructureInfo,
     ToolCallItem,
     _GetInfoFunc,
 )
@@ -29,7 +28,7 @@ def _safe_val(raw: str) -> Any:
             return raw
 
 
-class Qwen3XMLDetector(BaseFormatDetector):
+class Qwen3CoderDetector(BaseFormatDetector):
     """
     Detector for Qwen 3 models.
     Assumes function call format:
@@ -127,24 +126,26 @@ class Qwen3XMLDetector(BaseFormatDetector):
                 params[pname] = _safe_val(pval)
             raw = {"name": fname, "arguments": params}
             try:
+                # TODO: fix idx in function call, the index for a function
+                # call will always be -1 in parse_base_json
                 res.extend(self.parse_base_json(raw, tools))
             except Exception:
                 logger.warning("invalid tool call for %s dropped", fname)
         return res
 
+    def supports_structural_tag(self) -> bool:
+        return False
+
     def structure_info(self) -> _GetInfoFunc:
-        return lambda n: StructureInfo(
-            begin=f"{self.tool_call_start_token}\n<function={n}>",
-            end=f"</function>\n{self.tool_call_end_token}",
-            trigger=self.tool_call_start_token,
-        )
+        raise NotImplementedError
 
-    # TODO: fake ebnf for xml + outlines backend
     def build_ebnf(self, tools: List[Tool]):
         return EBNFComposer.build_ebnf(
             tools,
             individual_call_start_token=self.tool_call_start_token.replace("\n", "\\n"),
             individual_call_end_token=self.tool_call_end_token.replace("\n", "\\n"),
             tool_call_separator="\\n",
-            function_format="json",
+            function_format="xml",
+            call_rule_fmt='"<function={name}>\\n" {arguments_rule} "\\n</function>"',
+            key_value_rule_fmt='"<parameter={key}>\\n" {valrule} "\\n</parameter>"',
         )

sglang/srt/layers/activation.py

@@ -33,6 +33,7 @@ from sglang.srt.utils import (
     cpu_has_amx_support,
     is_cpu,
     is_cuda,
+    is_hip,
     is_npu,
     set_weight_attrs,
 )
@@ -42,9 +43,12 @@ _is_cuda = is_cuda()
 _is_npu = is_npu()
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
+_is_hip = is_hip()
 
 if _is_cuda:
     from sgl_kernel import gelu_and_mul, gelu_tanh_and_mul, silu_and_mul
+elif _is_hip:
+    from sgl_kernel import gelu_and_mul, gelu_quick, gelu_tanh_and_mul, silu_and_mul
 
 if is_npu():
     import torch_npu
@@ -126,9 +130,13 @@ class QuickGELU(CustomOp):
         return x * torch.sigmoid(1.702 * x)
 
     def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
-        # TODO(zhyncs): Implement the CUDA kernel for QuickGELU in sgl-kernel
         return self.forward_native(x)
 
+    def forward_hip(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty(x.shape, dtype=x.dtype, device=x.device)
+        gelu_quick(x, out)
+        return out
+
 
 class ScaledActivation(nn.Module):
     """An activation function with post-scale parameters.
@@ -222,8 +230,8 @@ def get_cross_encoder_activation_function(config: PretrainedConfig):
         return nn.Identity()
 
 
-if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available)):
+if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available) or _is_hip):
     logger.info(
-        "sgl-kernel is not available on Non-NV platforms or Non-AMX CPUs. Fallback to other kernel libraries."
+        "sgl-kernel is not available on Non-NV, Non-AMD platforms or Non-AMX CPUs. Fallback to other kernel libraries."
     )
     from vllm.model_executor.layers.activation import GeluAndMul, SiluAndMul

sglang/srt/layers/attention/base_attn_backend.py

@@ -65,7 +65,9 @@ class AttentionBackend(ABC):
         **kwargs,
     ):
         """Run forward on an attention layer."""
-        if forward_batch.forward_mode.is_decode():
+        if forward_batch.forward_mode.is_idle():
+            return q.new_empty(q.shape[0], layer.tp_q_head_num * layer.v_head_dim)
+        elif forward_batch.forward_mode.is_decode():
             return self.forward_decode(
                 q,
                 k,

sglang/srt/layers/communicator.py

@@ -24,8 +24,8 @@ from sglang.srt.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from sglang.srt.layers.dp_attention import (
-    attn_tp_all_gather,
-    attn_tp_reduce_scatter,
+    attn_tp_all_gather_into_tensor,
+    attn_tp_reduce_scatter_tensor,
     dp_gather_partial,
     dp_scatter,
     get_attention_dp_size,
@@ -309,8 +309,8 @@ class CommunicateSimpleFn:
             forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
             hidden_states,
         )
-        attn_tp_all_gather(
-            list(hidden_states.tensor_split(context.attn_tp_size)),
+        attn_tp_all_gather_into_tensor(
+            hidden_states,
             local_hidden_states,
         )
         return hidden_states
@@ -400,9 +400,7 @@ class CommunicateWithAllReduceAndLayerNormFn:
                 ].clone(),
                 residual,
             )
-            attn_tp_all_gather(
-                list(residual.tensor_split(context.attn_tp_size)), local_residual
-            )
+            attn_tp_all_gather_into_tensor(residual, local_residual)
         if context.attn_dp_size != 1:
             if context.attn_tp_rank == 0:
                 hidden_states += residual
@@ -442,9 +440,11 @@ class CommunicateWithAllReduceAndLayerNormFn:
         *,
         residual_input_mode,
     ):
-        tensor_list = list(hidden_states.tensor_split(context.attn_tp_size))
-        hidden_states = tensor_list[context.attn_tp_rank]
-        attn_tp_reduce_scatter(hidden_states, tensor_list)
+        input_hidden_states = hidden_states
+        hidden_states = hidden_states.tensor_split(context.attn_tp_size)[
+            context.attn_tp_rank
+        ]
+        attn_tp_reduce_scatter_tensor(hidden_states, input_hidden_states)
         if residual_input_mode == ScatterMode.TP_ATTN_FULL:
             residual = residual.tensor_split(context.attn_tp_size)[context.attn_tp_rank]
         if hidden_states.shape[0] != 0:
@@ -547,8 +547,8 @@ class CommunicateSummableTensorPairFn:
             forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
             hidden_states,
         )
-        attn_tp_all_gather(
-            list(hidden_states.tensor_split(context.attn_tp_size)),
+        attn_tp_all_gather_into_tensor(
+            hidden_states,
             local_hidden_states,
         )
         return hidden_states, residual

sglang/srt/layers/dp_attention.py

@@ -3,7 +3,8 @@ from __future__ import annotations
 import functools
 import logging
 from contextlib import contextmanager
-from typing import TYPE_CHECKING, List
+from enum import IntEnum, auto
+from typing import TYPE_CHECKING, List, Tuple
 
 import torch
 import triton
@@ -30,6 +31,34 @@ _LOCAL_ATTN_DP_SIZE = None
 _LOCAL_ATTN_DP_RANK = None
 
 
+class DPPaddingMode(IntEnum):
+
+    # Padding tokens to max length and then gather tokens using `all_gather_into_tensor`
+    MAX_LEN = auto()
+    # Padding tokens to sum length and then gather tokens using `all_reduce`
+    SUM_LEN = auto()
+
+    def is_max_len(self):
+        return self == DPPaddingMode.MAX_LEN
+
+    def is_sum_len(self):
+        return self == DPPaddingMode.SUM_LEN
+
+    @classmethod
+    def get_dp_padding_mode(cls, global_num_tokens: List[int]) -> DPPaddingMode:
+        # we choose the mode that minimizes the communication cost
+        max_len = max(global_num_tokens)
+        sum_len = sum(global_num_tokens)
+        if sum_len * 2 > max_len * get_attention_dp_size():
+            return cls.MAX_LEN
+        else:
+            return cls.SUM_LEN
+
+    @classmethod
+    def get_default_mode_in_cuda_graph(cls) -> DPPaddingMode:
+        return cls.MAX_LEN
+
+
 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
@@ -162,7 +191,7 @@ def disable_dp_size():
         _ATTN_DP_SIZE = old_dp_size
 
 
-def get_dp_local_info(forward_batch: ForwardBatch):
+def get_dp_local_info(forward_batch: ForwardBatch) -> Tuple[torch.Tensor, torch.Tensor]:
     # `get_dp_local_info` is only called in global DP gather and scatter. We use global DP rank here.
     dp_rank = get_attention_dp_rank()
 
@@ -221,7 +250,7 @@ def memcpy_triton(dst, src, dim, offset, sz, offset_src):
     memcpy_triton_kernel[grid](dst, src, offset, sz, offset_src, chunk_size, BLOCK_SIZE)
 
 
-def _dp_gather(
+def _dp_gather_via_all_reduce(
     global_tokens: torch.Tensor,
     local_tokens: torch.Tensor,
     forward_batch: ForwardBatch,
@@ -238,13 +267,6 @@ def _dp_gather(
             local_tokens.untyped_storage() is not global_tokens.untyped_storage()
         ), "aliasing between global_tokens and local_tokens not allowed"
 
-        # NOTE: During draft extend, the gathered_buffer is padded to num_tokens * (speculative_num_steps + 1).
-        # But the size of local_tokens is total accepted tokens. We need to reduce the local_num_tokens to the
-        # actual size of the accepted tokens.
-        if forward_batch.forward_mode.is_draft_extend():
-            shape_tensor = local_num_tokens.new_full((), local_tokens.shape[0])
-            local_num_tokens = torch.minimum(local_num_tokens, shape_tensor)
-
         memcpy_triton(
             global_tokens, local_tokens, 0, local_start_pos, local_num_tokens, False
         )
@@ -263,6 +285,38 @@ def _dp_gather(
         global_tokens[:] = tensor_model_parallel_all_reduce(global_tokens)
 
 
+def _dp_gather_via_all_gather(
+    global_tokens: torch.Tensor,
+    local_tokens: torch.Tensor,
+    forward_batch: ForwardBatch,
+    is_partial: bool,
+):
+    if not is_partial:
+        if get_attention_tp_rank() != 0:
+            local_tokens.fill_(0)
+    scattered_local_tokens = local_tokens.tensor_split(get_attention_tp_size())[
+        get_attention_tp_rank()
+    ]
+    get_attention_tp_group().reduce_scatter_tensor(scattered_local_tokens, local_tokens)
+    get_tp_group().all_gather_into_tensor(global_tokens, scattered_local_tokens)
+
+
+def _dp_gather(
+    global_tokens: torch.Tensor,
+    local_tokens: torch.Tensor,
+    forward_batch: ForwardBatch,
+    is_partial: bool,
+):
+    if forward_batch.dp_padding_mode.is_max_len():
+        _dp_gather_via_all_gather(
+            global_tokens, local_tokens, forward_batch, is_partial
+        )
+    else:
+        _dp_gather_via_all_reduce(
+            global_tokens, local_tokens, forward_batch, is_partial
+        )
+
+
 def dp_gather_partial(
     global_tokens: torch.Tensor,
     local_tokens: torch.Tensor,
@@ -296,24 +350,18 @@ def dp_scatter(
             local_tokens.untyped_storage() is not global_tokens.untyped_storage()
         ), "aliasing between local_tokens and global_tokens not allowed"
 
-        # NOTE: During draft extend, the gathered_buffer is padded to num_tokens * (speculative_num_steps + 1).
-        # But the size of local_tokens is total accepted tokens. We need to reduce the local_num_tokens to the
-        # actual size of the accepted tokens.
-        if forward_batch.forward_mode.is_draft_extend():
-            shape_tensor = local_num_tokens.new_full((), local_tokens.shape[0])
-            local_num_tokens = torch.minimum(local_num_tokens, shape_tensor)
-
         memcpy_triton(
             local_tokens, global_tokens, 0, local_start_pos, local_num_tokens, True
         )
 
 
-def attn_tp_reduce_scatter(
-    output: torch.Tensor,
-    input_list: List[torch.Tensor],
-):
-    return get_attention_tp_group().reduce_scatter(output, input_list)
+def attn_tp_reduce_scatter_tensor(output: torch.Tensor, input: torch.Tensor):
+    return get_attention_tp_group().reduce_scatter_tensor(output, input)
+
+
+def attn_tp_all_gather_into_tensor(output: torch.Tensor, input: torch.Tensor):
+    return get_attention_tp_group().all_gather_into_tensor(output, input)
 
 
-def attn_tp_all_gather(output_list: List[torch.Tensor], input_: torch.Tensor):
-    return get_attention_tp_group().all_gather(input_, output_tensor_list=output_list)
+def attn_tp_all_gather(output_list: List[torch.Tensor], input: torch.Tensor):
+    return get_attention_tp_group().all_gather(input, output_tensor_list=output_list)

sglang/srt/layers/logits_processor.py

@@ -27,7 +27,9 @@ from sglang.srt.distributed import (
     tensor_model_parallel_all_gather,
 )
 from sglang.srt.layers.dp_attention import (
+    DPPaddingMode,
     attn_tp_all_gather,
+    attn_tp_all_gather_into_tensor,
     dp_gather_replicate,
     dp_scatter,
     get_attention_dp_rank,
@@ -111,7 +113,8 @@ class LogitsMetadata:
     # Number of tokens to sample per DP rank
     global_num_tokens_for_logprob_cpu: Optional[torch.Tensor] = None
     global_num_tokens_for_logprob_gpu: Optional[torch.Tensor] = None
-
+    # The gather mode for DP attention
+    dp_padding_mode: Optional[DPPaddingMode] = None
     # for padding
     padded_static_len: int = -1
 
@@ -163,12 +166,12 @@ class LogitsMetadata:
             forward_batch_gathered_buffer=forward_batch.gathered_buffer,
             global_num_tokens_for_logprob_cpu=forward_batch.global_num_tokens_for_logprob_cpu,
             global_num_tokens_for_logprob_gpu=forward_batch.global_num_tokens_for_logprob_gpu,
+            dp_padding_mode=DPPaddingMode.SUM_LEN,
         )
 
-    def compute_dp_attention_metadata(self, hidden_states: torch.Tensor):
-        if self.global_num_tokens_for_logprob_cpu is None:
-            # we are capturing cuda graph
-            return
+    def compute_dp_attention_metadata(self):
+        # TODO(ch-wan): gathered_buffer here is larger than the actual required size in draft extend,
+        # we may use a smaller buffer in draft extend.
 
         cumtokens = torch.cumsum(self.global_num_tokens_for_logprob_gpu, dim=0)
         dp_rank = get_attention_dp_rank()
@@ -179,18 +182,9 @@ class LogitsMetadata:
         else:
             dp_local_start_pos = cumtokens[dp_rank - 1]
         dp_local_num_tokens = self.global_num_tokens_for_logprob_gpu[dp_rank]
-        gathered_buffer = torch.zeros(
-            (
-                sum(self.global_num_tokens_for_logprob_cpu),
-                hidden_states.shape[1],
-            ),
-            dtype=hidden_states.dtype,
-            device=hidden_states.device,
-        )
 
         self.dp_local_start_pos = dp_local_start_pos
         self.dp_local_num_tokens = dp_local_num_tokens
-        self.gathered_buffer = gathered_buffer
 
 
 class LogitsProcessor(nn.Module):
@@ -434,7 +428,7 @@ class LogitsProcessor(nn.Module):
         guarantee the given hidden_states follow this constraint.
         """
         if self.do_tensor_parallel_all_gather_dp_attn:
-            logits_metadata.compute_dp_attention_metadata(hidden_states)
+            logits_metadata.compute_dp_attention_metadata()
             hidden_states, local_hidden_states = (
                 torch.empty_like(logits_metadata.gathered_buffer),
                 hidden_states,
@@ -463,15 +457,31 @@ class LogitsProcessor(nn.Module):
 
         if self.do_tensor_parallel_all_gather:
             if self.use_attn_tp_group:
-                global_logits = torch.empty(
-                    (self.config.vocab_size, logits.shape[0]),
-                    device=logits.device,
-                    dtype=logits.dtype,
-                )
-                global_logits = global_logits.T
-                attn_tp_all_gather(
-                    list(global_logits.tensor_split(self.attn_tp_size, dim=-1)), logits
-                )
+                if self.config.vocab_size % self.attn_tp_size == 0:
+                    global_logits = torch.empty(
+                        (
+                            self.attn_tp_size,
+                            logits.shape[0],
+                            self.config.vocab_size // self.attn_tp_size,
+                        ),
+                        device=logits.device,
+                        dtype=logits.dtype,
+                    )
+                    attn_tp_all_gather_into_tensor(global_logits, logits)
+                    global_logits = global_logits.permute(1, 0, 2).reshape(
+                        logits.shape[0], self.config.vocab_size
+                    )
+                else:
+                    global_logits = torch.empty(
+                        (self.config.vocab_size, logits.shape[0]),
+                        device=logits.device,
+                        dtype=logits.dtype,
+                    )
+                    global_logits = global_logits.T
+                    attn_tp_all_gather(
+                        list(global_logits.tensor_split(self.attn_tp_size, dim=-1)),
+                        logits,
+                    )
                 logits = global_logits
             else:
                 logits = tensor_model_parallel_all_gather(logits)