fix(awq_sm70_moe): register int32 buffer Parameters with requires_grad=False

[jobordu]

What

vllm/model_executor/layers/quantization/awq_sm70_moe.py:_set_parameter (line 333 in v1.5.0; line 506 in current main) wraps a raw int32 buffer into an nn.Parameter without passing requires_grad=False to the constructor:

# before
param = value if isinstance(value, Parameter) else Parameter(value)
param.requires_grad_(False)

The constructor's default requires_grad=True triggers a UserWarning from PyTorch ≥ 2.x when value.dtype is torch.int32, because integer tensors cannot require gradients. The in-place requires_grad_(False) on the next line does silently turn the flag off, but only after the warning has already been emitted at boot.

# after
param = value if isinstance(value, Parameter) else Parameter(value, requires_grad=False)
param.requires_grad_(False)

The trailing param.requires_grad_(False) stays as a no-op idempotent guard for the already-Parameter branch.

Repro (was)

$ python -c "
import torch
import torch.nn as nn
from vllm.model_executor.layers.quantization.awq_sm70_moe import _set_parameter
layer = nn.Module()
_set_parameter(layer, 'qweight', torch.empty(4, dtype=torch.int32))
print('requires_grad =', layer.qweight.requires_grad)
"
UserWarning: Only Tensors of floating point and complex dtype can require gradients
requires_grad = False

After this PR

Same script runs silently; layer.qweight.requires_grad is still False. No other observable change for float tensors (the other six Parameter(...) call sites in process_weights_after_loading already pass requires_grad=False to the constructor for exactly this reason — this PR closes the same gap on the helper that those constructors' set_weight_attrs path does not delegate to).

Why fix this on _set_parameter and not on each call site

_set_parameter is the helper every AWQ SM_70 MoE weight writes flow through when a layer wants a raw tensor (no dtype cast) bound as a parameter. There are exactly six Parameter(torch.empty(...)) constructors explicit in process_weights_after_loading (lines 430/442/457/469/481/493 in v1.5.0, all already correct); any future call site that uses _set_parameter(...) for an int32 tensor inherits the fix for free. The fix is one line at the helper rather than N lines at every call site.

Scope

One file, one function, one line of behavioural change. No test added in this PR — stop there per the request that escalated this. The maintainers' preference on a regression test for "no UserWarning on int32 _set_parameter" can drive the next PR.

Out of scope

• The six explicit Parameter(torch.empty(...), requires_grad=False) sites are already correct. Left untouched.
• awq_marlin.py and awq_qpn_sm70.py are separate code paths; audit-and-fix if maintainers want a sweep.
• No SDK behaviour change for int8/float16/bfloat16 callers.

Environment

• 1Cat-vLLM v1.5.0 wheel on PyTorch 2.10.0+cu128.
• The bug is on the boot path of any AWQ SM_70 MoE expert that uses _set_parameter with an int32 buffer. The runtime workaround on the production side currently lives in commit bd9259a of Digital-Frontier-LDA/Flynn (a runtime sed against the installed wheel) — that workaround becomes unnecessary once this PR lands and a new tag is cut.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀