sgl-project/sglang · error · NotImplementedError

MiniMax H3 ring parallelism requires the FlashAttention back

Error message

MiniMax H3 ring parallelism requires the FlashAttention backend (matches --ring-degree's general restriction).

What it means

MiniMax H3's attention core supports ring (sequence) parallelism only via the FlashAttention backend; when ring is active and the configured backend differs it raises NotImplementedError, mirroring the global --ring-degree restriction.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py:597

            _usp_input_all_to_all_packed_qkv,
            _usp_output_all_to_all,
        )

        q, k, v = _usp_input_all_to_all_packed_qkv(q, k, v)

    if attention._attention_impl is None:
        attention._set_attention_backend(
            get_attn_backend(
                attention.head_dim,
                q.dtype,
                attention_requirements=AttentionRequirements(packed_varlen=True),
            )
        )

    if ring_active:
        ring_ws, _ = get_ring_ctx()
        if attention._attention_backend_enum is not AttentionBackendEnum.FA:
            raise NotImplementedError(
                "MiniMax H3 ring parallelism requires the FlashAttention "
                "backend (matches --ring-degree's general restriction)."
            )
        # max_seqlen is cu_seqlens[1] (`used`) by construction -- the real,
        # non-padding row count ring needs, already a host int here.
        out = _ring_attention_varlen(
            q,
            k,
            v,
            attn_impl=attention._attention_impl,
            real_seq_len=max_seqlen,
            ring_ws=ring_ws,
        )
    else:
        if (
            attention._attention_backend_enum
            is AttentionBackendEnum.SUBBLOCK_SPARSE_ATTN
        ):

View on GitHub (pinned to 0132848349)

Solutions

1. Set the attention backend to FA when using ring parallelism (e.g. --attention-backend fa / SGLANG_ATTN_BACKEND default) and keep --ring-degree
2. Drop ring parallelism (ring-degree 1 / remove the flag) if you must use the non-FA backend
3. Verify with get_ring_ctx() / server args at startup that ring and backend settings are consistent before serving

Example fix

# before
python -m sglang.launch_server --model minimax-h3 --ring-degree 2 --attention-backend flashinfer

# after
python -m sglang.launch_server --model minimax-h3 --ring-degree 2 --attention-backend fa

from sglang.srt.layers.attention import AttentionBackendEnum
assert attention_backend == AttentionBackendEnum.FA or ring_degree == 1, \
    'ring parallelism requires FA backend'

When it happens

Trigger: Launching with --ring-degree > 1 (ring_active true) while the attention backend is set to something other than FA (e.g. FlashInfer, Triton, Torch) in server args or model runner config.

Common situations: Combining sequence parallelism with a backend chosen for other models on the same deployment; backend override flags or env vars; upgrading sglang where default backend changed away from FA.

Related errors

• MiniMax-H3 ring parallelism requires the FlashAttention back
• AITer backend does not have a metadata builder.
• AITER Sage backend does not have a metadata builder.
• {type(self).__name__} does not implement packed varlen atten
• {type(self).__name__} does not implement ring KV-chunk atten

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/efad5183209aa635. Report an issue: GitHub.