Expand Troubleshooting with HF auth, FA2 cu_seqlens, and smoke tests#79
Open
lonexreb wants to merge 1 commit intoNVlabs:mainfrom
Open
Expand Troubleshooting with HF auth, FA2 cu_seqlens, and smoke tests#79lonexreb wants to merge 1 commit intoNVlabs:mainfrom
lonexreb wants to merge 1 commit intoNVlabs:mainfrom
Conversation
Address recurring questions in the issue tracker by extending the existing Troubleshooting section: 1. New entry for the IndexError that physical_ai_av raises when the gated PhysicalAI-Autonomous-Vehicles dataset is inaccessible — reproduces the exact traceback fragment users will see and links to the auth section. Addresses NVlabs#59, NVlabs#61. 2. New entry for the FA2 cu_seqlens_q crash in the diffusion expert path when FA2 is forced globally — explains the 4D-mask root cause in one paragraph and the shipped VLM-FA2 / expert-SDPA split. Addresses NVlabs#52. 3. New entry with three import-level smoke tests (torch.cuda, alpamayo_r1 import, huggingface-cli whoami) plus the end-to-end test_inference.py command, so users can rapidly bisect a broken environment. Addresses NVlabs#33. No code changes. Signed-off-by: lonexreb <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Doc-only expansion of the existing Troubleshooting section in
README.md. No code changes. Addresses several recurring issues by giving users a fast path to a fix without filing a new ticket.What's added
1.
IndexError: list index out of rangewhen loading the datasetReproduces the exact traceback fragment users see when HF auth or gated-dataset access is missing, and points to the existing §3 Authenticate with HuggingFace step. Pairs with PR #74 which surfaces the same diagnostic at runtime.
→ Addresses #59, #61.
2.
RuntimeError: cu_seqlens_q must have shape (batch_size + 1)with FlashAttention-2One-paragraph explanation: the diffusion expert uses a 4D additive attention mask to handle variable-length VLM rollouts, which FA2 cannot consume. The shipped configuration keeps the VLM on FA2 and the expert on SDPA — so the error only appears when users force FA2 globally. Mentions the
expert_cfg.attn_implementation = "sdpa"escape hatch.→ Addresses #52. Pairs with PR #75 which makes this default explicit in code.
3. Verifying your local installation
Three quick import-level smoke tests (
torch.cuda.is_available(),from alpamayo_r1.models.alpamayo_r1 import AlpamayoR1,huggingface-cli whoami) plus the end-to-endtest_inference.pycommand, so users can bisect a broken environment in a few seconds.→ Addresses #33.
Why this is worth doing as docs
Looking at the open-issue distribution: the same handful of root causes (HF auth, FA2 incompatibility with the expert, missing GPU/imports) explain a large fraction of bug reports. A focused Troubleshooting section is the smallest-blast-radius fix that also reduces maintainer triage load.
Test plan
#3-authenticate-with-huggingfaceresolve to the existing heading.