skill(infer-hw-adapt, infer-precision-check): expand backend coverage and add shared precision check skill

[physics31415926]

Summary

Two skill updates based on findings from recent backend adaptation work.

1. infer-hw-adapt: expand backend coverage (MUSA, Sunrise, Iluvatar, Hygon, PPU, MetaX), fix duplicate content. Version numbers are runtime-detected at Stage 0, not hardcoded.
2. infer-precision-check: new shared skill providing a standardized correctness gate for inference. Usable from model porting, hardware adaptation, plugin version upgrades, and regression detection.

Changes

infer-hw-adapt

SKILL.md

• Add 5 new backend keywords: musa, iluvatar, sunrise, hygon, ppu
• Expand breakage table with 7 new patterns (platform detection, patch idempotency, torch.accelerator stub, ops registry conflict, CUDAGraph timeout, FlagGems fast-path guard, MoE dispatch registration)
• Fix duplicate content: original file contained two full copies of Stage 0–6; deduplicated to one
• Improve Stage 4 serving test: add curl request command and monitor parameters
• Simplify Stage 0 probe command to essential checks

SUMMARY.md

• Add platform detection to the Load when description
• Fix constraint rule count: 14 → 7 (aligns with actual Critical Rules list)

infer-precision-check (new skill)

SKILL.md

• 3-tier comparison protocol: Tier 1 exact token ID → Tier 2 top-K set → Tier 3 semantic
• 7-step workflow: env probe → prompts → GT collection (NVIDIA) → target collection → compare → multimodal → TP scaling → serving
• Reusable Python scripts: gt_collect.py, target_collect.py, compare.py
• Failure diagnosis guide: mismatch at token 0–2 / 3–10 / 10+ / systematic failure
• Pass/fail checklist and PR result recording template
• 5 constraint guards: greedy_only, gt_first, same_model_weights, log_all_outputs, tp_parity

SUMMARY.md

• Concise card covering load-when triggers, full cycle, key principles, and tier definitions

Integration

• infer-model-adapt calls infer-precision-check at Step 13 (after offline inference passes)
• infer-hw-adapt calls infer-precision-check at Stage 4 (after functional tests pass)

No functional changes to agent behavior

All updates are documentation/guidance only. No code changes.

---

Precision Verification: Qwen3.6-27B on MetaX C550

Real-world validation run using the infer-precision-check skill on MetaX C550 hardware.

Setup

Item	GT (NVIDIA A800)	Target (MetaX C550)
Hardware	A800-SXM4-80GB	MetaX C550
Container	vllm-precision-gt	vllm-fl-adapt-0202
vLLM	0.20.2 (CUDA)	0.20.2+empty + vllm-plugin-fl 0.2.0
Model	Qwen3.6-27B (shared weights, same files)	Qwen3.6-27B (same)
Decoding	greedy (temperature=0)	greedy (temperature=0)
TP	1	1
max_tokens	128	128

Tier 1 Results (Exact Token ID Match)

5 prompts × 128 max tokens, greedy decoding:

#	Prompt	GT len	MX len	Result
0	What is the capital of France?	128	128	❌ MISMATCH @ token 35
1	Explain the theory of relativity in simple terms.	128	128	❌ MISMATCH @ token 108
2	Write a Python function to compute fibonacci numbers.	128	128	✅ MATCH
3	What are the main differences between TCP and UDP?	128	128	✅ MATCH
4	Translate Hello how are you into Chinese.	11	11	✅ MATCH

Tier 1: 3/5 MATCH

Analysis

• Prompts 2, 3, 4 achieve exact token-level match — basic compute path is correct.
• Prompts 0 and 1 diverge late (token 35 and token 108 respectively), consistent with BF16 floating-point accumulation order differences between CUDA and MACA backends causing different greedy selections at numerically close logit positions. This is expected behavior for non-identical hardware.
• Divergence is not at token 0–2 (would indicate weight loading or tokenizer mismatch) — weights and tokenizer are confirmed identical.
• Tier 2 (top-K logit set match) verification was attempted but blocked by GPU memory contention on the A800 node during the test window. To be completed in a follow-up run.

Known Workaround Applied

MetaX MACA NCCL init_process_group (world_size=1) consumes ~54 GiB of device memory before vLLM's memory snapshot, causing the default gpu_memory_utilization=0.85 check to fail (free=7.54 GiB < requested=54 GiB). Fix: use kv_cache_memory_bytes=3*1024**3 to bypass the utilization-based check and directly specify KV cache size. Also requires enforce_eager=True.