Skip to content

Replay Phase 2 and Phase 6 T-C5 live 8 GiB RTX 3080 telemetry on geminus #4

Open

Description

@ahrazzle

Phase 2 hardware discovery and Phase 6 T-C5 serving-budget measurements in Proven-Commons/mcproof PR 3 executed on a local 64 GiB Apple M1 Max (llama-server 0.5.0, $0 cloud) because the geminus 8 GiB RTX 3080 host (ASUS UX582HS) had Tailscale SSH and Linux workbench ports closed (linux_rtx3080_online = false).

This issue tracks replaying live nvidia-smi peak VRAM and tok/s telemetry on geminus under an active Tailscale SSH session.

Analytical Baseline Already Verified

From compile/artifacts/phase2_hardware_checkpoint_report.json and verification/VERIFICATION_RECEIPT.json under siliconcordat/mde issue 4:

Staged Checkpoint SHA-256 Linux Headless VRAM (8k F16 KV + 3 F16 LoRAs) Windows WDDM VRAM Ceiling (7.2 GiB)
Qwen3.5-4B-Q8_0.gguf 10cc391b403021dd11c614679d2fd92f611c3681d29e29651b717316965d61e1 4.9451 GiB 6.1451 GiB PASS
Qwen3.5-4B-Q4_K_M.gguf 00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4 3.4034 GiB 4.6034 GiB PASS

Replay Checklist on geminus

  1. Confirm Tailscale SSH connectivity to geminus and verify nvidia-smi + llama-server 0.5.0+ (--lora-scaled, --lora-init-without-apply, --ctx-size 8192).
  2. Stage Qwen3.5-4B-Q8_0.gguf (10cc391b...) and Qwen3.5-4B-Q4_K_M.gguf (00fe7986...) alongside the three F16 rank-8 LoRAs in compile/artifacts/composition/ (corpus, weft-eido, formal-verification).
  3. Run compile/src/mcproof_compile/hardware_probe.py and MCPROOF_LIVE_LLAMA_TESTS=1 python3 -m pytest stack/tests/test_stack_suite.py -v.
  4. Record live nvidia-smi peak allocated VRAM (<= 7.2 GiB requirement), 8-slot 8192-token KV cache residency, and single-slot / 8-slot tok/s overhead (<= 15.0% vs pure W0) in stack/results/final/phase6_receipt.json and docs/CONTINUITY.md.

Automated posting by agentic team with human oversight.

Activity

  1. ahrazzle commented on Nov 19, 2026

    @ahrazzle
    MemberAuthor

    Live Phase 6 T-C5 8 GiB RTX 3080 telemetry replay on the ASUS UX582HS workbench (Windows 11 Build 26200, NVIDIA GeForce RTX 3080 Laptop GPU 8192 MiB, driver 566.07, CUDA 12.7, llama-server b12549) and 3-OS cross-ISA (arm64 macOS, x86_64 Linux, and AMD64 Windows 11) CPU float64 re-derivation via weft.cross_isa (siliconcordat/weft PR 4, commit 7176902419ec11b9454225934a75736b62dd3b47) are both complete.

    1. Live 8 GiB RTX 3080 T-C5 Replay (45528746556e4236dcb4a8d10d6fed5b26673c57)

    Both Qwen3.5-4B checkpoints were launched with -c 65536 -np 8 --cache-type-k f16 --cache-type-v f16 --lora-init-without-apply and all three F16 rank-8 LoRAs (corpus d81bab6b..., weft-eido 71bedfee..., formal-verification 728272c0...), polling nvidia-smi every 100 ms across 5 interleaved W0 vs triple-LoRA trials plus an 8-slot concurrent 7486-token/slot KV residency stress test:

    Checkpoint (SHA-256) Idle WDDM VRAM Peak nvidia-smi VRAM (<= 7.2 GiB) 8-Slot 8192 KV Residency 1-Slot W0 -> Stack (<= 15.0%) 8-Slot W0 -> Stack (<= 15.0%)
    Qwen3.5-4B-Q8_0.gguf (10cc391b...) 0 MiB 6.8525 GiB (7017 MiB) PASS 8 x 8192 (7486/slot live) PASS 64.16 -> 63.91 tok/s (+0.38%) PASS 282.78 -> 284.65 tok/s (-0.66%) PASS
    Qwen3.5-4B-Q4_K_M.gguf (00fe7986...) 0 MiB 5.2314 GiB (5357 MiB) PASS 8 x 8192 (7486/slot live) PASS 86.27 -> 85.86 tok/s (+0.47%) PASS 346.55 -> 344.89 tok/s (+0.48%) PASS

    Live measurement also isolated two offsetting modeling assumptions in hardware_probe.py: NVIDIA Optimus drives both UX582HS panels on the Intel iGPU so idle Windows WDDM dGPU usage is 0 MiB (rather than the +1.20 GiB allowance), while -np 8 -c 65536 (kv_unified=false) allocates 8 x 8192 F16 KV slots (~2.0 GiB plus 8x recurrent state) rather than 1 x 8192 (0.25 GiB), leaving 0.35 GiB of Q8_0 headroom below the 7.2 GiB gate and returning to 0 MiB after shutdown.

    2. Three-OS Cross-ISA (arm64 vs x86_64) BetaEdit Re-Derivation

    Re-derivation across arm64 macOS (Apple Accelerate, Python 3.12.15, NumPy 2.5.5), x86_64 Linux (Netlib CBLAS/LAPACK 3.12.0, Python 3.14.8, NumPy 2.5.4), and AMD64 Windows 11 (Python 3.12.10, NumPy 2.5.5) pinned the legacy Weft cross-ISA drift boundary and confirmed 0-byte parity under weft.cross_isa:

    Adapter / Suite Legacy Weft (arm64 vs x86_64) weft.cross_isa (macOS arm64 / Linux x86_64 / Win11 AMD64)
    formal-verification P2-B (f16 / int4) PASS (f16 7ffa8780..., cert 56820652...) PASS (0-byte match: f16 e2caa16f..., cert cb571814...)
    formal-verification P2-B (float64 / f32) FAIL (float64 diff 5.816e-13, 5 byte diffs in f32) PASS (0-byte match: f32 dd04281f..., cert 498bbd38...)
    weft-eido (eido_soft / eido_hard f16) FAIL (8bed53d4... vs 0b01806c..., 9898c603... vs 17995dd0...) PASS (0-byte match: e0dbf794... / 12970dcc...)
    PCG64(20261119) 3-OS portable suite Diverges across BLAS/LAPACK backends PASS (0-byte match across all 3 OS/ISA hosts in < 1 s)

    What Is Not Proved Here

    • T-C5 ran via a standalone stdlib replica of the phase6_composition.py T-C5 harness (python -m unittest tests.test_stack_suite: 10 passed, 3 lab-only skipped). Running the full sibling-repo T-C1..T-C4 pytest suite (mde, exo, warp_modeld, weft) natively on UX582HS awaits WSL2/Linux toolchain bring-up.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions