sgl-project/sglang · error · ValueError

Incomplete Diffusers H3 fused parameters

Error message

Incomplete Diffusers H3 fused parameters: {incomplete}

What it means

When loading a MiniMax H3 checkpoint in Diffusers fused format, the loader tracks a set of expected fused parameters; if some remain unfilled after processing (pending), the fused state dict is incomplete and it raises listing the missing names. This guards against silently partial weight loads.

Solutions

  1. Inspect the 'incomplete' names in the message and compare against the checkpoint's actual state_dict keys to see the naming mismatch
  2. Re-export or re-download the checkpoint with a matching Diffusers version, or use the native (non-Diffusers) weight layout
  3. Update the name-mapping table in _diffusers_h3_checkpoint to the new key names, mapping each pending tensor

Example fix

# before: checkpoint uses 'blocks.0.attn.qkv.weight' but map expects fused names
state = torch.load('h3_diffusers.pt')

# after: rename to expected fused layout or load native names
for k in list(state):
    state[k.replace('attn.qkv', 'attn.fused_qkv')] = state.pop(k)
Defensive patterns

Strategy: validation

Validate before calling

expected = set(EXPECTED_FUSED_NAMES)
actual = set(state_dict.keys())
missing = expected - actual
assert not missing, f'checkpoint missing fused keys: {sorted(missing)}'

Try / catch

try: load_diffusers_h3(sd)\nexcept ValueError as e: fallback_to_native_layout(sd) if 'Incomplete' in str(e) else raise

Prevention

When it happens

Trigger: Loading a Diffusers-format H3 checkpoint whose keys don't cover all expected fused parameters — renamed keys across Diffusers versions, a partially exported/sliced checkpoint, or a version skew between the loader's expected name map and the checkpoint.

Common situations: Upgrading Diffusers or sglang where fused qkv/mlp parameter names changed; using a community-converted or re-exported checkpoint missing some fused tensors; the test test_native_weight_names_and_grouped_qkv_reorder exercising name mapping with a stale fixture.

Related errors


AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/09c06ecddc33376a. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py:125

        if merge_index is None:
            yield target_name, tensor
            continue

        assert merge_count is not None
        pending[target_name][merge_index] = tensor
        if len(pending[target_name]) != merge_count:
            continue

        merge_dim = 1 if target_name.endswith((".qweight", ".qzeros", ".scales")) else 0
        yield target_name, torch.cat(
            [pending[target_name][index] for index in range(merge_count)],
            dim=merge_dim,
        )
        del pending[target_name]

    if pending:
        incomplete = ", ".join(sorted(pending))
        raise ValueError(f"Incomplete Diffusers H3 fused parameters: {incomplete}")


_BF16_DTYPE = torch.bfloat16
_FP32_DTYPE = torch.float32
_MPS_MLP_TOKEN_CHUNK_SIZE = 128
# keep MPS activation chunks below the allocator high-watermark; CUDA keeps
# its fused full-sequence projection
_MPS_QKV_PROJECTION_TOKEN_CHUNK_SIZE = 128
_MPS_ATTENTION_QUERY_TOKEN_CHUNK_SIZE = 128

_MPS_EMBED_WEIGHT_PREFIXES = (
    "condition_proj",
    "video_patch_proj",
    "audio_patch_proj",
    "time_embedder",
    "token_refiner.final_norm",
)

View on GitHub (pinned to 0132848349)