[jason-fxz]

Summary

Loading a quantized checkpoint prints the kernel choice once per layer: select_kernel runs for every layer's QuantMethod and logs each time. nvidia/Qwen3.6-35B-A3B-NVFP4 (ModelOpt mixed precision, W4A16_NVFP4 experts) prints 121 lines with only 2 distinct messages:

kernel triton selected; skipped w4a4: the experts declare no static activation scale (W4A16 checkpoint)
kernel triton selected; skipped b12x: the weight declares no static activation scale (W4A16 checkpoint); cublas: the weight declares no static activation scale (W4A16 checkpoint)

This adds debug_once, info_once and warning_once to the loggers init_logger returns, following vLLM's logger.info_once (default scope "local"): rank 0 only, once per distinct (logger, level, msg, args) in the process. The seen set is module-level because init_logger rebuilds a logger's methods on every call for the same name. Args must hash by value, as with vLLM's lru_cache key.

Moved onto it:

• layers/quantization/method.py: select_kernel.
• layers/quantization/moe/nvfp4.py: the "installed but unusable" warnings of the marlin and b12x symbol checks, which unusable_reason runs for every MoE layer.
• layers/quantization/linear/nvfp4.py: _warn_part_globals (functools.cache + warning_rank0) removed.
• models/gguf/mmproj.py: _warn_inline_vision (functools.cache per model path) removed.
• models/minimax_m3/config.py: _log_mode_once and _LOGGED_MODES removed.
• models/qwen3_vl/vision.py: the _warned_rope_fallback flag removed.

These lines now log on rank 0 only; before, the minimax, mmproj, qwen3_vl and MoE symbol-check lines logged on every rank.

Left as they are, since they suppress differently: scheduler.py _warned_cut_image (first warning only, the text varies per image), server/request_logger.py _warn_once (one flag for every message), moe/offload_cache.py hit-d2d flag (re-arms after a rebuild).

Tests

• 1x RTX PRO 6000 Blackwell Server Edition (sm_120, 96 GB), driver 595.91.07, torch 2.14.1+cu130.
• ft serve --model nvidia/Qwen3.6-35B-A3B-NVFP4 --port 30112 (local copy): 121 kernel ... selected lines on main 9cacd08, 2 on this branch.
• pytest -m "not slow" tests/models/test_quant_config.py tests/models/test_minimax_m3.py tests/models/test_minimax_m3_vision.py tests/models/test_qwen3_vl.py tests/models/test_qwen3_vl_vision.py tests/models/qwen4_exp/test_skeleton.py tests/models/test_gemma4_gguf_rope.py tests/models/test_models_registry.py tests/kernels/test_nvfp4_w4a4_linear.py tests/kernels/test_e4m3_compat.py tests/moe: 393 passed, 50 skipped, 8 failed; the same 8 fail on main 9cacd08 (7 in test_quant_config.py from local checkpoints whose quant dialect differs from the HF repo the case names, or index-only MiniMax-M3-NVFP4 snapshots; 1 in tests/moe/test_nvfp4_backends.py::test_w4a4_prefill_matches_nvfp4_activation_reference[512-1024-swiglu_clamp-256], which passed in an earlier run on both).
• No new test.