sgl-project/sglang · error · NotImplementedError

The 'enable-mixed-chunk' feature is currently unsupported in

Error message

The 'enable-mixed-chunk' feature is currently unsupported in the following scenarios: 1. When using the MLA backend on Ascend NPU devices, 2. When using the deepseekv3.2 model on Ascend NPU devices, 3. When the environment variable ASCEND_USE_FIA is set to 0 and qk_head_dim exceeds 128 on Ascend NPU devices.

What it means

forward_mixed implements mixed-chunk (combined prefill+decode) attention on the Ascend backend, but only for the plain, non-MLA, FIA-enabled path. It raises NotImplementedError when topk_indices are supplied (sparse/MTP routing), when the MLA backend is in use (covers deepseekv3.2 MLA on Ascend), or when FIA is off and qk_head_dim > 128.

Source

Thrown at python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py:2907

    def forward_mixed(
        self,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        layer: RadixAttention,
        forward_batch: ForwardBatch,
        save_kv_cache: bool = True,
        q_rope: Optional[torch.Tensor] = None,
        k_rope: Optional[torch.Tensor] = None,
        topk_indices: Optional[torch.Tensor] = None,
    ):
        if (
            topk_indices is not None
            or self.use_mla
            or (not self.use_fia and layer.qk_head_dim > 128)
        ):
            raise NotImplementedError(
                "The 'enable-mixed-chunk' feature is currently unsupported in the following scenarios: "
                "1. When using the MLA backend on Ascend NPU devices, "
                "2. When using the deepseekv3.2 model on Ascend NPU devices, "
                "3. When the environment variable ASCEND_USE_FIA is set to 0 and qk_head_dim exceeds 128 on Ascend NPU devices."
            )
        if save_kv_cache:
            self.token_to_kv_pool.set_kv_buffer(
                layer,
                KVWriteLoc(
                    forward_batch.out_cache_loc,
                    self.forward_metadata.swa_out_cache_loc,
                ),
                k,
                v,
            )
        k_cache = self.token_to_kv_pool.get_key_buffer(layer.layer_id)
        v_cache = self.token_to_kv_pool.get_value_buffer(layer.layer_id)
        num_block, block_size, _, _ = k_cache.shape

View on GitHub (pinned to 0132848349)

Solutions

1. Remove --enable-mixed-chunk from server args on Ascend for these model configs
2. Set ASCEND_USE_FIA=1 if your only blocker is the qk_head_dim>128 non-FIA clause
3. Switch to a non-MLA model or disable speculative/MTP features that pass topk_indices
4. Track the SGLang Ascend roadmap for mixed-chunk MLA support

Example fix

# before
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2 --enable-mixed-chunk
# after
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2

def mixed_chunk_ok_on_ascend(use_mla, use_fia, qk_head_dim, topk_used):
    return not use_mla and (use_fia or qk_head_dim <= 128) and not topk_used

def can_use_mixed_chunk_ascend(backend, layer) -> bool:
    return (
        not backend.use_mla
        and (backend.use_fia or layer.qk_head_dim <= 128)
    )

try:
    backend.forward_mixed(...)
except NotImplementedError as e:
    if "enable-mixed-chunk" in str(e):
        # fall back to separate prefill/decode scheduling
        run_unmixed_schedule()

When it happens

Trigger: Launching with --enable-mixed-chunk on Ascend plus any of: MLA backend active (e.g. deepseek models with MLA), deepseekv3.2, topk_indices (speculative/MoE routing) passed to the mixed path, or ASCEND_USE_FIA=0 with a layer whose qk_head_dim > 128.

Common situations: Copying a CUDA-oriented launch config with --enable-mixed-chunk to an Ascend deployment of DeepSeek-family models; speculative decoding feeding topk indices into mixed batches.

Related errors

• CP attention for non-FIA path on Ascend is not yet implement
• NPU packed attention does not support a sequence that is emp
• Ascend PD transfer does not support HiSparse destination dev
• Short-conv hybrid models (ZAYA1 CCA, LFM2 / LFM2-MoE) are no
• {name} is required for NPU packed attention

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/930ad09c5079160b. Report an issue: GitHub.