docs(ocrbench): the one-batch image decode is a per-host fact, not the build - #330
Merged
Merged
Conversation
…e build "GGUF arms therefore decode the image in one batch (ADR 0036)" holds on the discrete CUDA card the CUDA ladder was measured on. It does not hold on gfx1151, where ADR 0036's own diagnostic fires on every gemma4 load and the 2048 floor is refused at every quantisation -- because availableMemoryForLoad takes its integrated-GPU branch and sizes the batch against 31 GiB of system RAM while the scheduler logs 95.4 GiB of GPU free. Left as-is the row invites the next reader to infer a decode path from a build stamp on whatever host they are on. Points at the ROCm doc, which carries the measurement and the arithmetic; ADR 0036 already records it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
glennneuber
pushed a commit
that referenced
this pull request
Sep 19, 2026
#330 landed the same correction with the root cause named — availableMemoryForLoad takes its integrated-GPU branch on gfx1151 and sizes the batch against 31 GiB of system RAM while the scheduler logs 95.4 GiB of GPU. That is the better text. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The last piece of #326 that #327 did not already absorb.
The CUDA ladder's config table says:
True on that discrete card. False on gfx1151, where ADR 0036's own diagnostic fires on every gemma4 load:
The 2048 floor is refused at every quantisation — 1024 for q4 at
-np 1, 512 for q8 and bf16 — becauseavailableMemoryForLoadtakes its integrated-GPU branch and sizes the batch against 31 GiB of system RAM while the scheduler logsavailable="95.4 GiB"of GPU. On a discrete card that branch never triggers.As written the row invites the next reader to infer a decode path from a build stamp regardless of host. Now qualified, and pointing at the ROCm doc that carries the measurement.
🤖 Generated with Claude Code