sgl-project/sglang · error · ValueError

DeepSeekV4 CP supports moe_a2a_backend in {supported_a2a_bac

Error message

DeepSeekV4 CP supports moe_a2a_backend in {supported_a2a_backends}, got {cfg.moe_a2a_backend!r}.

What it means

When DeepSeek V4 runs with prefill context parallelism, the MoE all-to-all backend must be one of none, deepep, megamoe, or mori because only these backends have correct collective/precision behavior under CP for this model. Any other moe_a2a_backend is rejected.

Source

Thrown at python/sglang/srt/arg_groups/deepseek_v4_hook.py:209

    declare_resolution(
        server_args,
        "validate_deepseek_v4_cp",
        moe_dense_tp_size=1,
    )
    declare_resolution(
        server_args,
        "validate_deepseek_v4_cp",
        attn_cp_size=cfg.tp_size // cfg.dp_size,
    )
    assert (
        cfg.dp_size == 1
    ), "For round-robin split mode, dp attention is not supported."
    assert (
        cfg.tp_size <= 8
    ), "Context parallel only supports single machine (tp_size <= 8). Cross-machine CP has precision issues."
    supported_a2a_backends = ("none", "deepep", "megamoe", "mori")
    if cfg.moe_a2a_backend not in supported_a2a_backends:
        raise ValueError(
            f"DeepSeekV4 CP supports moe_a2a_backend in {supported_a2a_backends}, "
            f"got {cfg.moe_a2a_backend!r}."
        )
    logger.warning(
        "Disabling SGLANG_OPT_FLASHMLA_SPARSE_PREFILL because DeepSeekV4 "
        "context parallelism is enabled."
    )
    envs.SGLANG_OPT_FLASHMLA_SPARSE_PREFILL.set(False)
    logger.warning(
        f"Enable Context Parallel for DeepSeekV4, "
        f"dp_size={cfg.dp_size}, moe_dense_tp_size={cfg.moe_dense_tp_size}, "
        f"attn_cp_size={cfg.attn_cp_size}, ep_size={cfg.ep_size}, tp_size={cfg.tp_size}"
    )

View on GitHub (pinned to 0132848349)

Solutions

1. Set --moe-a2a-backend to one of none, deepep, megamoe, or mori
2. Remove the explicit --moe-a2a-backend flag to use the default
3. If you need a specific backend, disable --enable-prefill-cp

Example fix

# before
--enable-prefill-cp --moe-a2a-backend some_backend
# after
--enable-prefill-cp --moe-a2a-backend deepep

SUPPORTED = {"none", "deepep", "megamoe", "mori"}
if getattr(cfg, 'enable_prefill_cp', False) and cfg.moe_a2a_backend not in SUPPORTED:
    server_args.moe_a2a_backend = 'deepep'

except ValueError as e:
    if 'moe_a2a_backend' in str(e): server_args.moe_a2a_backend = 'deepep'
    else: raise

When it happens

Trigger: DeepSeek V4 + enable_prefill_cp with cfg.moe_a2a_backend outside ("none", "deepep", "megamoe", "mori") — e.g. "flashinfer_tkv" or another a2a backend — during _handle_model_specific_adjustments.

Common situations: Reusing an a2a backend tuned for a different model/cluster, or a launcher template that sets --moe-a2a-backend unconditionally; also defaulting to a backend not yet ported to CP paths.

Related errors

• DeepSeekV4 only supports interleave CP strategy, got {cfg.cp
• DSpark with dp attention supports moe_a2a_backend 'none' (bu
• DSpark ignores --speculative-moe-a2a-backend; with dp attent
• DSA indexer weights_proj LoRA is incompatible with piecewise
• DSA indexer only supports CUDA, HIP, and NPU

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/393f0aae4681fa03. Report an issue: GitHub.