feat(preset): add GPT-OSS 120B presets for H100 SXM

[gangwon]

Summary

• Add 12 vLLM v0.17.0 presets for openai/gpt-oss-120b on NVIDIA H100 SXM
• Covers E2E, prefill, and decode roles with tp4/tp8 parallelism
• Includes both EP and TP MoE variants for each configuration

Test plan

• helm lint passes for moai-inference-preset
• helm template renders all 12 new presets correctly
• Verify preset naming and labels match conventions

[hhk7734]

Code review

Found 2 issues (each affects all 12 new preset files):

1. Missing --no-enable-log-requests in ISVC_EXTRA_ARGS (deploy/helm/AGENTS.md says "Logging arguments (--disable-uvicorn-access-log, --no-enable-log-requests) — presets must include these because ISVC_EXTRA_ARGS in a preset fully overrides the runtime base's value during Odin strategic merge patch"). The existing GPT-OSS 120B NVL preset includes it at line 32; PR MAF-19524: feat(preset): add vLLM v0.17.0 E2E presets for AI& April launch models #101 fixed the same omission across earlier presets.

mif/deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp4-moe-ep4.helm.yaml

Lines 27 to 38 in 730fd29

	- name: ISVC_EXTRA_ARGS
	value: >-
	--enable-auto-tool-choice
	--tool-call-parser openai
	--reasoning-parser openai_gptoss
	--max-num-seqs 128
	--max-num-batched-tokens 8192
	--max-cudagraph-capture-size 2048
	--max-model-len -1
	--disable-uvicorn-access-log
	--exclude-tools-when-tool-choice-none
	resources:

2. Missing --trust-remote-code in ISVC_EXTRA_ARGS. openai/gpt-oss-120b loads custom remote code; without this flag vLLM fails to load the model. Every other GPT-OSS preset sets it (e.g. NVL preset at line 28), and deploy/helm/AGENTS.md lists --trust-remote-code among the model-specific args presets own.

mif/deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp4-moe-ep4.helm.yaml

Lines 26 to 38 in 730fd29

	env:
	- name: ISVC_EXTRA_ARGS
	value: >-
	--enable-auto-tool-choice
	--tool-call-parser openai
	--reasoning-parser openai_gptoss
	--max-num-seqs 128
	--max-num-batched-tokens 8192
	--max-cudagraph-capture-size 2048
	--max-model-len -1
	--disable-uvicorn-access-log
	--exclude-tools-when-tool-choice-none
	resources:

Apply the same fix to the other 11 new files under deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[Copilot]

Pull request overview

Adds new Odin InferenceServiceTemplate preset manifests to support running openai/gpt-oss-120b on NVIDIA H100 SXM with vLLM v0.17.0, covering E2E, prefill, and decode roles across TP4/TP8 and MoE TP/EP variants.

Changes:

• Added 12 new vLLM v0.17.0 presets for openai/gpt-oss-120b on nvidia/h100-sxm.
• Included E2E, prefill, and decode roles for tp4 and tp8 parallelism.
• Provided both MoE TP (-moe-tp*) and MoE EP (-moe-ep*) variants for each role/TP size.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-prefill-nvidia-h100-sxm-tp8-moe-tp8.helm.yaml	New prefill preset (TP8, MoE-TP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-prefill-nvidia-h100-sxm-tp8-moe-ep8.helm.yaml	New prefill preset (TP8, MoE-EP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-prefill-nvidia-h100-sxm-tp4-moe-tp4.helm.yaml	New prefill preset (TP4, MoE-TP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-prefill-nvidia-h100-sxm-tp4-moe-ep4.helm.yaml	New prefill preset (TP4, MoE-EP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp8-moe-tp8.helm.yaml	New E2E preset (TP8, MoE-TP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp8-moe-ep8.helm.yaml	New E2E preset (TP8, MoE-EP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp4-moe-tp4.helm.yaml	New E2E preset (TP4, MoE-TP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp4-moe-ep4.helm.yaml	New E2E preset (TP4, MoE-EP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-decode-nvidia-h100-sxm-tp8-moe-tp8.helm.yaml	New decode preset (TP8, MoE-TP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-decode-nvidia-h100-sxm-tp8-moe-ep8.helm.yaml	New decode preset (TP8, MoE-EP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-decode-nvidia-h100-sxm-tp4-moe-tp4.helm.yaml	New decode preset (TP4, MoE-TP) for H100 SXM
deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-decode-nvidia-h100-sxm-tp4-moe-ep4.helm.yaml	New decode preset (TP4, MoE-EP) for H100 SXM

[gangwon]

GPT-OSS 120B는 사실 trust-remote-code가 없어도 돌긴 하는데 일단 있어도 문제는 안 되어서 리뷰를 따랐습니다 ㅎㅎ

[hhk7734]

Both issues addressed in 516931a — verified all 12 files now include --trust-remote-code and --no-enable-log-requests.

mif/deploy/helm/moai-inference-preset/templates/presets/vllm/v0.17.0/openai-gpt-oss-120b-nvidia-h100-sxm-tp4-moe-ep4.helm.yaml

Lines 27 to 39 in 516931a

	- name: ISVC_EXTRA_ARGS
	value: >-
	--trust-remote-code
	--enable-auto-tool-choice
	--tool-call-parser openai
	--reasoning-parser openai_gptoss
	--max-num-seqs 128
	--max-num-batched-tokens 8192
	--max-cudagraph-capture-size 2048
	--max-model-len -1
	--disable-uvicorn-access-log
	--no-enable-log-requests
	--exclude-tools-when-tool-choice-none

🤖 Generated with Claude Code