You can deploy this model and start earning money today!

pirola/
vllm-patched-calib

Name: vllm-patched-calib
Availability: InStock

Views

No views yet

vllm-patched-calib

vLLM 0.21.0 (commit ad7125a) with calibration-v2 hooks patch applied.

In-graph cudagraph-safe capture (this build): every capture signal now accumulates via pure GPU tensor ops recorded INTO the captured graph (the prior Python _ch.dispatch callback path yielded zero under cudagraphs, the production default). Traced-region signals (imatrix dense / block_out / layer_in) use named-arg mutates_args custom ops in calibration_custom_ops.py so Inductor cannot DCE them; custom-op-interior signals accumulate inline in MoERunner / TritonExperts. All 10 writers keep public captured_entry_count() for the driver fail-fast and dump from GPU accumulators (checkpoint schema v2). Runs cudagraph-fast (NOT enforce_eager).

Source repo: https://github.com/lucaspirola/moe_compress (branch feat/b0-hook-fix)
Patch artifacts (also uploaded to this repo for traceability):
- vllm_calibration_hooks.patch — 10846 lines, MD5 c19ffcc5d03fec7dc20caebfad80d7c1
- vllm_calibration_stage2_profile.patch — 941 lines, MD5 fca4c01cc5cef188b30c323bb919cd72
Architectures: sm_80 (A100), sm_90a (H100/H200), sm_100 (B200), sm_120 (RTX 6000 Pro Blackwell)
Build host: HF Jobs (cpu-performance)
torch: 2.11.0+cu130
CUDA toolkit: 13.0

Install on a fresh GPU host

bash

hf download pirola/vllm-patched-calib --include "*.whl" --local-dir /tmp/wheels
pip install /tmp/wheels/vllm-*.whl

Calibration capture flags

The patched vLLM accepts new env vars to enable calibration data capture:

VLLM_CALIB_CAPTURE_ROUTER=1 — per-layer router logits + topk
VLLM_CALIB_CAPTURE_EXPERT=1 — per-expert inputs + weighted outputs
VLLM_CALIB_CAPTURE_EXPERT_UNWEIGHTED=1 — kernel-level pre-weight per-expert outputs (Triton backend; forces VLLM_USE_FLASHINFER_MOE_FP16=0)
VLLM_CALIB_CAPTURE_EXPERT_MID=1 — silu(gate)·up intermediate (input to down_proj; Triton backend)
VLLM_CALIB_CAPTURE_BLOCK=1 — MoE block pre-residual output
VLLM_CALIB_CAPTURE_IMATRIX=1 — per-input-channel sum-of-squares for every linear layer (writes llama.cpp-compatible .imatrix.dat)
VLLM_CALIB_CAPTURE_INPUT_COV=1 — per-(layer, expert, "gate_proj") teacher input covariance Σ_in (requires VLLM_CALIB_CAPTURE_EXPERT=1; writes dict-shaped sidecars/covariance.pt, schema v2)
VLLM_CALIB_MAX_LAYER=<N> — L2 early-exit gate: truncate Qwen3MoeModel.forward after decoder layer N (inclusive); skips N+1..end. Default -1 / unset = disabled. Orthogonal to the capture gates above. Useful as the foundation for L1 (sequential REAP+REAM per-layer profiling) and as a standalone optimisation for any writer whose payload comes from layer L or earlier.

Spot-preemption resumability

The driver build_self_traces_calib_vllm.py writes a periodic <jsonl>.imatrix.ckpt checkpoint at every chunk boundary (CLI: --imatrix-checkpoint-every-chunks=1 by default). On --resume, the checkpoint is hydrated into the live accumulators in-place and the cumulative prompt counter is restored. The final .imatrix.dat and the periodic .imatrix.ckpt both use the temp-file + os.replace atomic-rename pattern so a kill mid-write leaves the previous file intact. .npz logit sidecars are also written atomically.

JSONL resume is hardened against trailing partial lines: each line is JSON-validated on resume; the first parse failure triggers a truncate to the last good byte offset before counting resumes.