Repository navigation
train: refuse only a node NO scheduler backend can run (fix the Metal regression from #35) - #38
Merged
joelteply merged 1 commit intoOct 6, 2026
Conversation
… regression from #35) #35 refused a training graph whenever the DEVICE backend said no. Metal has no GET_ROWS_BACK kernel and has always run it on the CPU, so on the M5 every /train was refused ("the training graph has a node MTL0 cannot run: GET_ROWS_BACK ...", Fable, 2026-10-06). I verified #35 only on CUDA. The refusal now applies only when no backend of the scheduler supports the node. A device "no" with another backend's "yes" falls back as before. supports_op is truthful since #35, so a "yes" is a kernel that will not abort: the original crash (CUDA claiming a viewed SQR it would assert on) now falls back instead. Measured on the 5090, 0.8B qwen35, with the L2_NORM rule fix TEMPORARILY reverted so the viewed SQR is really in the graph: the job TRAINS through the CPU fallback (loss 0.5909) and /health stays ok. No abort, no refusal. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The regression: #35 refused a training graph whenever the device backend said no. Metal has no
GET_ROWS_BACKkernel and has always run it on the CPU, so on the M5 every/trainwas refused ("the training graph has a node MTL0 cannot run: GET_ROWS_BACK …", Fable, 2026-10-06). I verified #35 only on CUDA. My miss.The fix:
supports_opis truthful since train: a node the device cannot run refuses the JOB, not the server; supports_op tells the truth #35, so a "yes" is a kernel that won't abort. The original crash (CUDA claiming a viewed SQR it would assert on) now falls back instead.Measured on the 5090, 0.8B qwen35, with the L2_NORM rule fix temporarily reverted so the viewed SQR is really in the graph: the job trains through the CPU fallback (loss 0.5909), and
/healthstays ok. No abort, no refusal.Needs a Metal run before pin (Fable):
GET_ROWS_BACKmust fall back to the CPU and train.🤖 Generated with Claude Code
https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc