[ENV-STOP] Execution host fails frozen GPU identity gate (RTX 4090 Laptop 16 GB / Windows); #56/#57/#71/#72/#73 NOT executed

[HainingYin]

Summary

This host was assigned to run the OpenLithoHub 1×RTX 4090 GPU Execution Handoff (Track A: #56 → #57; Track B: #71 → #72 → #73). On pre-flight inspection the host does not satisfy the frozen hardware gates. Per the handoff GPU identity gate (§8) and STOP rules (§14), execution STOPPED before any clone, preflight, benchmark, or source/parameter change.

Per the handoff: "GPU 不是 RTX 4090 → STOP；不要执行 #72 / #73".

All five target issues are NOT EXECUTED on this host. Nothing was fabricated; results below are directly observed hardware/OS facts, not benchmark output.

Gate-by-gate verdict

Gate	Handoff requirement	Observed on this host	Verdict
GPU model (§4 / §8 identity gate)	NVIDIA GeForce RTX 4090 (desktop, 24 GB)	NVIDIA GeForce RTX 4090 Laptop GPU (16 GB)	FAIL
VRAM (§4)	24 GB	15.99 GB	FAIL
OS (§4)	Linux	Windows 11 (10.0.26200)	FAIL
nvidia-smi (§16 evidence)	required	Failed to initialize NVML: Unknown Error	FAIL
CUDA PyTorch	CUDA-enabled	torch 2.6.0+cu124, CUDA 12.4, cuDNN 9.1, cc 8.9	OK

The GPU is a different SKU (laptop/mobile 16 GB variant, not the desktop 24 GB), and the OS is Windows, not Linux. Both are frozen acceptance criteria, so the host is non-qualifying.

Environment evidence (directly observed, torch-reported)

• OS: Windows-10-10.0.26200-SP0 (Windows 11), MINGW64
• Python: 3.10.11
• PyTorch: 2.6.0+cu124 (CUDA 12.4)
• cuDNN: 9.1 (90100)
• torch.cuda.is_available: True
• device_count: 1
• GPU name: NVIDIA GeForce RTX 4090 Laptop GPU
• VRAM: 15.99 GB
• SM count: 76, compute capability 8.9 (Ada)
• nvidia-smi / nvidia-smi -q / nvidia-smi topo -m: all fail with Failed to initialize NVML: Unknown Error (NVML unavailable → driver version not readable)

Frozen commits (both verified reachable in OpenLithoHub/OpenLithoHub)

• Track A (V2): a70d4ac76fc73e1f1e1099c7884db9bfdd4802d3 (2026-09-25) — "Merge pull request [PR-G Phase 2B.2] Close remaining Industrial Benchmark v2 measurement-authority gaps before CUDA handoff #59 from OpenLithoHub/pr-g-phase2b2-authority-repair"
• Track B (Scale / RTX 4090): 210ed1b3a647303dfa00c43cd90a063184a0a831 (2026-09-26) — "Merge pull request [PR-S6] Migrate Industrial Scale track to 1×RTX4090 + Lane C saturation #70 from OpenLithoHub/pr-s6-scale-4090-migration" (current main HEAD)

Per-issue status on this host

Issue	Track	Frozen commit	Status	Reason
#56 CUDA Host Qualification	A	a70d4ac	NOT EXECUTED	Host fails frozen GPU identity / VRAM / OS gates (§8/§14)
#57 Formal Industrial Benchmark v2	A	a70d4ac	NOT EXECUTED	Blocked by #56 + env gate
#71 G1-4090 Scale Host Qualification	B	210ed1b	NOT EXECUTED	Host fails GPU identity gate (not desktop RTX 4090 24 GB)
#72 G2-4090 Large-Layout Single-GPU Scale	B	210ed1b	NOT EXECUTED	Handoff §8: "GPU 不是 RTX 4090 → 不要执行 #72/#73"
#73 G3-4090 Single-GPU Microbatch Saturation	B	210ed1b	NOT EXECUTED	Handoff §8: "GPU 不是 RTX 4090 → 不要执行 #72/#73"

Superseded / deferred (NOT executed, per handoff §2): #66, #67, #68, #69.

What was / was not done

Done

• Verified repo reachability, both frozen commits, and the five target issues (all present, open).
• Captured direct hardware / OS / CUDA / PyTorch / cuDNN evidence.
• Confirmed nvidia-smi/NVML is non-functional on this host.

Not done (by design, per STOP gate)

• Did NOT clone the repo, run any preflight, run any benchmark, modify source, modify parameters, run the §16 evidence archive, or execute GPU Issue G1 — Scale Host Qualification (3×RTX 3080) #66/GPU Issue G2 — Large-Layout Single-GPU Scale (Lane A) #67/GPU Issue G3 — 1/2/3-GPU Scaling (Lane B) #68/GPU Issue G4 — Optional Hopkins Compute Stress (P2, bounded) #69.
• Because execution stopped at the hardware identity gate, the §16 fields run identity, verifier result, archive SHA-256, final clean-tree state, and driver are N/A (not reached) — they are deliberately absent, not omitted.

To proceed, a qualifying host is required

1 × NVIDIA GeForce RTX 4090   (desktop, 24 GB VRAM)
Linux
CUDA-enabled PyTorch + cuDNN
KLayout
working nvidia-smi (NVML)

Re-run from the top of the handoff on that host: #56 → #57 (a70d4ac) and #71 → #72 → #73 (210ed1b).

Cross-references

#56 #57 #71 #72 #73

[HainingYin]

Update: superseded by the Laptop16GB handoff — execution proceeded on this host

The STOP recorded in this issue was based on the desktop-24GB generation of the handoff,
whose hardware gate required a desktop RTX 4090 (24 GB) + Linux. A revised handoff —
OpenLithoHub_1xRTX4090_Laptop16GB_GPU_Execution_Handoff.md — was then issued for exactly
this host: it explicitly qualifies the RTX 4090 Laptop GPU (16 GB) as the target (both
desktop and Laptop variants satisfy the identity gate; VRAM ≈ 16 GB is the expected class),
and its operational gates are the preflight results plus the GPU identity check — not the OS
or the 24 GB class.

Per that handoff, execution was carried out on this host (2026-09-26). Current state:

Issue	Result
#56 (Track A host qualification)	FAIL — preflight PASS 16/16, but the frozen v2 harness at a70d4ac has two source defects (Tier B: reset_peak_memory_stats before CUDA init; Tier C: HopkinsParams(sigma_outer=…) kwarg mismatch within the commit). Full diagnosis posted to #56 (comment 5845405820).
#57 (Track A formal)	NOT EXECUTED — blocked by #56 FAIL (a source fix would invalidate the freeze).
#71 (Track B scale qualification)	PASS — preflight PASS 16/16, GPU identity gate PASS, bounded Lane C smoke SUCCESS (provisional). Reported on #71 (comment 5846183628).
#72 (Track B large-layout scale)	IN PROGRESS — 2/4 rungs verified SUCCESS (4096 median 74.80 s; 8192 median 283.55 s); 16384/32768 running in a detached run (ETA 2026-09-27 ≈ 10:50). Reported on #75.
#73 (Track B microbatch saturation)	QUEUED — chained after #72 (ETA 2026-09-27 ≈ 13:00). Reported on #75.
#66/#67/#68/#69	NOT EXECUTED (superseded/deferred per handoff §2).

Consolidated Track B report (environment facts, verified rung numbers, interruption +
resume transparency, accommodations, follow-up plan): #75.

Host environment notes that remain true and matter for maintainers:

• nvidia-smi/NVML is non-functional on this host (Failed to initialize NVML: Unknown Error); all GPU facts are torch-reported (GPU model/VRAM/CC all correct).
• The frozen codebase has multiple Linux-only assumptions that block or complicate execution
  on this Windows host: hardcoded /usr/bin/git (Track A scripts and the scale harness
  — the scale harness's git_measurement_commit calls /usr/bin/git directly), POSIX-only
  import resource (scale harness), CRLF vs LF manifest regeneration, --provisional flag
  in issue text not present in the frozen CLI. Each is documented with evidence in the
  respective issue comments and in the operator's execution report.
• Every result from this host is labeled Windows 11 — NOT frozen-protocol-compliant;
  reference-only, per the handoff.

This issue is kept as the historical record of the gen-1 gate evaluation.