sgl-project/sglang · error · ValueError

Module instance {} is not unique

Error message

Module instance {} is not unique 

What it means

register_hooks found the same module instance already present in module_to_name_map, i.e. the same nn.Module object was encountered twice (typically because it is shared/reused and named_modules() yields it again under another name, or hooks were registered twice).

Source

Thrown at python/sglang/srt/utils/nvtx_pytorch_hooks.py:291

        skip_types = (
            torch.nn.Identity,
            torch.nn.Dropout,
            torch.nn.Dropout1d,
            torch.nn.Dropout2d,
            torch.nn.Dropout3d,
        )

        for name, module in network_model.named_modules(prefix=module_prefix):
            # Skip certain module types to reduce profiling overhead
            if isinstance(module, skip_types):
                continue

            module.register_forward_pre_hook(self.module_fwd_pre_hook)
            module.register_forward_hook(self.module_fwd_hook)
            if module not in self.module_to_name_map:
                self.module_to_name_map[module] = name
            else:
                raise ValueError("Module instance {} is not unique ".format(module))
        return

View on GitHub (pinned to 0132848349)

Solutions

1. Deduplicate by id(module) before registering (skip already-seen instances)
2. Remove existing hooks before re-registering (track and remove_forward_hook handles)
3. Use named_modules(remove_duplicate=True) (the default) and avoid memo overrides

Example fix

# before
for name, module in model.named_modules():
    self._register_one(name, module)
# after
seen = set()
for name, module in model.named_modules():
    if id(module) in seen:
        continue
    seen.add(id(module))
    self._register_one(name, module)

seen = set()
unique_modules = [(n, m) for n, m in model.named_modules() if not (id(m) in seen or seen.add(id(m)))]

def is_unique_module(module, registered: set) -> bool:
    return id(module) not in registered

try:
    hooker.register_hooks(model)
except ValueError as e:
    if 'not unique' in str(e):
        logger.warning('shared modules present; skip NVTX hooks')

When it happens

Trigger: Calling register_hooks on a model with weight-tied/shared modules where named_modules() revisits the same instance, or calling register_hooks twice without unregistering.

Common situations: NVTX profiling instrumentation on models with tied embeddings (lm_head shared with embed_tokens), or re-registering hooks across profiling sessions.

Related errors

• No trace files found for profile_id: {self.profile_id}
• manually start is only supported yet
• manually stop is only supported yet
• unsupported profile stage: {forward_mode=}
• Trace file is empty.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/def8ad6bf0ed79f2. Report an issue: GitHub.