Skip to content

[NV] H100 (Agg): migrate model path #1537

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Ankur-singh wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[NV] H100 (Agg): migrate model path #1537

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ce884ce
h100(agg): migrate model path to /mnt/numa1/shared/models/, drop hf d…
Ankur-singh May 20, 2026
3ab2f5f
perf-changelog: set PR link to #1537
Ankur-singh May 20, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 4 additions & 26 deletions .github/configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4504,15 +4504,7 @@ gptoss-fp4-h100-vllm:
- isl: 1024
osl: 1024
search-space:
- { tp: 2, conc-start: 4, conc-end: 64 }
- { tp: 4, conc-start: 4, conc-end: 64 }
- { tp: 8, conc-start: 4, conc-end: 64 }
- isl: 8192
osl: 1024
search-space:
- { tp: 2, conc-start: 4, conc-end: 64 }
- { tp: 4, conc-start: 4, conc-end: 64 }
- { tp: 8, conc-start: 4, conc-end: 16 }
- { tp: 2, conc-start: 4, conc-end: 4 }

minimaxm2.5-fp8-h100-vllm:
image: vllm/vllm-openai:v0.21.0
Expand All @@ -4527,13 +4519,7 @@ minimaxm2.5-fp8-h100-vllm:
- isl: 1024
osl: 1024
search-space:
# - { tp: 8, ep: 8, conc-start: 4, conc-end: 64 }
- { tp: 4, ep: 4, conc-start: 4, conc-end: 64 }
- isl: 8192
osl: 1024
search-space:
# - { tp: 8, ep: 8, conc-start: 4, conc-end: 64 }
- { tp: 4, ep: 4, conc-start: 4, conc-end: 64 }
- { tp: 4, ep: 4, conc-start: 4, conc-end: 4 }

# Diverged from minimaxm2.5-fp8-h100-vllm (agentic-coding sibling). Metadata is
# identical to origin/main's minimaxm2.5-fp8-h100-vllm; the split exists because this
Expand Down Expand Up @@ -9210,11 +9196,7 @@ qwen3.5-fp8-h100-sglang:
- isl: 1024
osl: 1024
search-space:
- { tp: 8, ep: 8, conc-start: 4, conc-end: 32 }
- isl: 8192
osl: 1024
search-space:
- { tp: 8, ep: 8, conc-start: 4, conc-end: 32 }
- { tp: 8, ep: 8, conc-start: 4, conc-end: 4 }

qwen3.5-fp8-h100-sglang-mtp:
image: lmsysorg/sglang:v0.5.12-cu130
Expand All @@ -9229,8 +9211,4 @@ qwen3.5-fp8-h100-sglang-mtp:
- isl: 1024
osl: 1024
search-space:
- { tp: 8, ep: 8, conc-start: 4, conc-end: 32, spec-decoding: mtp }
- isl: 8192
osl: 1024
search-space:
- { tp: 8, ep: 8, conc-start: 4, conc-end: 32, spec-decoding: mtp }
- { tp: 8, ep: 8, conc-start: 4, conc-end: 4, spec-decoding: mtp }
1 change: 0 additions & 1 deletion benchmarks/single_node/gptoss_fp4_h100.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,6 @@ if [[ -n "$SLURM_JOB_ID" ]]; then
echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
fi

if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi

MAX_MODEL_LEN=10240

Expand Down
1 change: 0 additions & 1 deletion benchmarks/single_node/minimaxm2.5_fp8_h100.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,6 @@ if [[ -n "$SLURM_JOB_ID" ]]; then
echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
fi

if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi

nvidia-smi

Expand Down
1 change: 0 additions & 1 deletion benchmarks/single_node/qwen3.5_fp8_h100.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,6 @@ fi

nvidia-smi

if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
Expand Down
1 change: 0 additions & 1 deletion benchmarks/single_node/qwen3.5_fp8_h100_mtp.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,6 @@ fi

nvidia-smi

if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi

export SGLANG_ENABLE_SPEC_V2=1

Expand Down
11 changes: 11 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3050,3 +3050,14 @@
description:
- "Update SGLang image from v0.5.11-cu130 (5d old) to v0.5.12-cu130"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1475

- config-keys:
- gptoss-fp4-h100-vllm
- minimaxm2.5-fp8-h100-vllm
- qwen3.5-fp8-h100-sglang
- qwen3.5-fp8-h100-sglang-mtp
description:
- "Migrate H100 agg model path to /mnt/numa1/shared/models/ (audit-recommended root); set HF_HUB_CACHE to container's ~/.cache/huggingface for eval dataset downloads"
- "Rewrite MODEL in launcher from HF id (org/name) to absolute local path under HF_HUB_CACHE_MOUNT; remove hf download from bench scripts so contract is staged-or-fail"
- "Reduce search-space to single (isl=1024, osl=1024, conc=4) point per config to verify model-path wiring end-to-end"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX

Check failure on line 3063 in perf-changelog.yaml

View check run for this annotation

Claude / Claude Code Review

perf-changelog: pr-link placeholder XXXX not replaced 

The new perf-changelog entry at line 3063 has `pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX` — the literal placeholder `XXXX` was never replaced with the real PR number (1537). Every other recent entry in the file uses a real PR id, and per AGENTS.md this file is the audit trail linking config changes back to the PR that triggered them. Please replace `XXXX` with `1537` before merging.
Copy link
Copy Markdown
Contributor

@claude claude Bot May 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 The new perf-changelog entry at line 3063 has pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX — the literal placeholder XXXX was never replaced with the real PR number (1537). Every other recent entry in the file uses a real PR id, and per AGENTS.md this file is the audit trail linking config changes back to the PR that triggered them. Please replace XXXX with 1537 before merging.

Extended reasoning...

What the bug is. perf-changelog.yaml line 3063 reads:

  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX

XXXX is a forgotten-substitution placeholder, not a real PR id. The author clearly intended to fill in a real number — the rest of the new entry (config-keys list and three-line description) is complete and substantive — but the pr-link field was left with its template value.

Why this matters. Every other recent entry in the same file uses a real PR id, e.g.:

  • line 3037 → pull/1411
  • line 3044 → pull/1528
  • line 3052 → pull/1475

AGENTS.md describes perf-changelog.yaml as the chronological audit trail linking benchmark config changes back to the PR that triggered them. An unfilled XXXX permanently breaks that linkage for the four H100 agg configs touched by this change (gptoss-fp4-h100-vllm, minimaxm2.5-fp8-h100-vllm, qwen3.5-fp8-h100-sglang, qwen3.5-fp8-h100-sglang-mtp) once merged — anyone later trying to understand why those four configs were narrowed to a single (isl=1024, osl=1024, conc=4) point will hit a dead 404 link instead of this PR.

Why nothing prevents it. There is no pre-commit or CI check that validates perf-changelog.yaml pr-link URLs resolve to a real PR; it relies on author discipline. The file is structured YAML but XXXX is still a syntactically valid string, so no schema validator will flag it either.

Impact. Once merged, the link https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX will 404 forever (GitHub does not have a PR numbered XXXX). The audit trail for these four H100 configs is severed. It is also non-trivial to fix post-merge because (a) perf-changelog.yaml is append-only by convention, and (b) reviewers in 6 months won't know to look for XXXX specifically.

Step-by-step proof.

  1. Look at the diff hunk for perf-changelog.yaml: the new entry ends at line 3063 with pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX.
  2. The PR number is 1537 (from the PR metadata at the top of this review).
  3. Visit https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX — GitHub will return 404 because no PR is numbered XXXX.
  4. Compare with line 3052: pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1475 — that one resolves to a real PR. So the convention in this file is clearly to use a real PR id.

How to fix. One-character (well, four-character) substitution at perf-changelog.yaml:3063:

-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1537

16 changes: 14 additions & 2 deletions runners/launch_h100-dgxc-slurm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,7 +279,19 @@ EOF

else

HF_HUB_CACHE_MOUNT="/mnt/nfs/sa-shared/gharunners/hf-hub-cache/"
HF_HUB_CACHE_MOUNT="/mnt/numa1/shared/models/"

# HF_HUB_CACHE is set to help with dataset download inside the container
# for eval jobs. Can be updated to some other path on the cluster and
# mounted just like HF_HUB_CACHE_MOUNT.
export HF_HUB_CACHE="$HOME/.cache/huggingface"

# Rewrite MODEL from HF id (org/name) to the pre-staged local path under
# HF_HUB_CACHE_MOUNT. Skip if MODEL is already an absolute path.
if [[ -n "$MODEL" && "$MODEL" != /* ]]; then
export MODEL="${HF_HUB_CACHE_MOUNT}${MODEL##*/}"
fi

SQUASH_FILE="/mnt/nfs/lustre/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
LOCK_FILE="${SQUASH_FILE}.lock"

Expand All @@ -306,7 +318,7 @@ else

srun --jobid=$JOB_ID \
--container-image=$SQUASH_FILE \
--container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
--container-mounts=$GITHUB_WORKSPACE:/workspace/,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE_MOUNT \
--no-container-mount-home \
--container-workdir=/workspace/ \
--no-container-entrypoint --export=ALL,PORT=8888 \
Expand Down