Skip to content

docs(inference): show how hardware profiles identify the GPU - #14760

Merged
cv merged 1 commit into
v1from
fix/spark-gb10-identification
Nov 21, 2026
Merged

cv merged 1 commit into
v1from
fix/spark-gb10-identification

Conversation

@cv

@cv cv commented Nov 21, 2026

Copy link
Copy Markdown
Collaborator

Documents that managed-service hardware profiles identify the GPU by the name nvidia-smi reports, shows how to check that name before applying, and lists DGX Spark systems whose driver reports a placeholder name as a current limit.

Problem

On a DGX Spark whose driver reports NVIDIA JMJWOA-Generic-GPU, the GTC DC Hermes example plans, then apply fails after creating the gateway and model container:

service runtime stopped: GPU-…: hardware.profile requires NVIDIA GB10; observed different GPU family

The docs described the requirement as an ARM64 GB10 without saying that the runtime identifies the GPU by its nvidia-smi name, so a reader couldn't check a host before applying.

Part of #14698: these docs say that such a host can't use the profile. Whether v1 should accept another GPU identity source remains open there.

Change

  • docs/agents-inference/models.md: how each profile matches the GPU name, the failure for a placeholder name, and the nvidia-smi command to run before applying.
  • examples/gtc-dc/README.md: the same check in the host preparation steps.
  • docs/resources/limits.md: a row for DGX Spark systems that report a placeholder GPU name.

The runtime's check is unchanged. Accepting the placeholder would trust a name that the driver emits for a GPU it doesn't identify; NemoClaw's main line also sees the same placeholder from the Windows-on-Arm WSL shim on non-NVIDIA hardware. Identifying GB10 by PCI device ID or firmware product name needs data from an affected host first, requested on #14698.

Beyond CI

  • On a DGX Spark with driver 580.178.04, the documented command printed NVIDIA GB10, 12.1, 580.178.04 on the host, and the same name appeared inside a --gpus all container.
  • Passing the reported placeholder line through the runtime's nvidia-smi decoder and the dgx-spark check produced the documented error; the same line with NVIDIA GB10 passed.
  • The same Spark ran a managed vLLM service in October 2026 on driver 580.142, the reporter's version. This suggests that the driver version alone doesn't produce the placeholder.
  • Not verified: the affected host's PCI device ID, and whether a driver update changes its reported name.

@cv

cv commented Nov 21, 2026

Copy link
Copy Markdown
Collaborator Author
Authoring session — opening
Field Status Value Evidence or limit
Harness recorded Pi 1.6.0 (session format v3) Session header and the harness install path in the system prompt; cutoff 2026-11-21T22:52:58Z
Skills recorded verify-yaml-deployment (complete read, with targets/dgx-spark.md); nemoclaw-contributor-create-pr (complete read) Tool results in the session record to cutoff
Instruction files recorded AGENTS.md (injected); WRITING.md (complete read); .github/PULL_REQUEST_TEMPLATE.md (complete read); .github/SESSION_REPORT_TEMPLATE.md (complete read) System prompt injection and tool results to cutoff
Tool calls partial typescript: 63 Wrapper calls, including failed ones, to cutoff; nested shell, git, gh, and workspace calls inside each program are not recorded separately
Tokens recorded input: 7622065; cached input: 7320656; output: 52663 Sum of 65 unique assistant responses' usage (input + cacheRead + cacheWrite; cacheRead; output) to cutoff; all responses reported nonzero counters
Session window recorded 2026-11-21T22:18Z to 2026-11-21T22:53Z; 34 min First record to cutoff, including idle time
Compactions recorded 0 Complete record to cutoff
Models recorded claude-opus-5-5 Model change record and every response; no changes
Reasoning levels recorded high Thinking-level record and every response; no changes

@cv
cv enabled auto-merge (squash) November 21, 2026 22:53
@github-actions github-actions Bot added the v1 NemoClaw v1 branch label Nov 21, 2026
@github-actions

Copy link
Copy Markdown
Contributor

@github-actions

Copy link
Copy Markdown
Contributor

Selected hardware E2Es

Commit: d68fa55991507ccfed34ba2a0f19937a10618700

Selected cases run automatically; results do not gate merging.

Shared code or test infrastructure changed.

  • dgx-station-openclaw-nemotron
  • docker-openclaw-hermes-ollama
  • docker-openclaw-llama-cpp
  • docker-openclaw-vllm
  • docker-openclaw-vllm-shared

@cv
cv merged commit 051fa65 into v1 Nov 21, 2026
100 checks passed
@cv
cv deleted the fix/spark-gb10-identification branch November 21, 2026 23:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

v1 NemoClaw v1 branch

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant