xnnpack: size and describe an fp32 export as fp32 - #9
Merged
Merged
Conversation
…s carried the 8da4w recipe and sizing The fp32 recipe (dcd6288) changed what export_llm builds but not what the report says about it. Every SmolLM2-360M-Instruct fp32 report on the Hub claims int4_codes "round-to-nearest", group size 32, "4-bit weights in groups of 32" in its description, and the 8da4w sizing: 1.33 GiB resident at 8k for a file of 1.40 GiB. execubench's model scan caught it. - sizing: bytes per linear weight per recipe (8da4w 0.5625, fp32 4); pte_bytes_estimate, device_resident_bytes and choose_context take the qmode. SmolLM2-360M fp32 at 2k: 1,496,317,952 B from the terms against 1,496,334,720 B measured. - export: the recipe is settled before sizing, so the window table, device_resident_bytes and pte_bytes_estimate describe the file built; recipe_fields writes int4_codes and group_size as null for fp32, with its own label and description. 8da4w reports are unchanged. - plan --qmode sizes the phone estimate for either recipe. Co-Authored-By: Claude Opus 5.5 <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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 fp32 recipe (dcd6288) changed what
export_llmbuilds but not what the report says about it. Every SmolLM2-360M-Instruct fp32 report on the Hub claimsint4_codes: round-to-nearest, group size 32, "4-bit weights in groups of 32" in its description, and the 8da4w sizing (1.33 GiB resident at 8k for a 1.40 GiB file). execubench's model scan caught it.pte_bytes_estimate,device_resident_bytesandchoose_contexttake the qmode. SmolLM2-360M fp32 at 2k: 1,496,317,952 B from the terms vs 1,496,334,720 B measured.recipe_fieldswritesint4_codesandgroup_sizeas null for fp32, with its own label and description. 8da4w reports are unchanged.--qmodesizes the phone estimate for either recipe.Tests: 207 passed, 4 skipped. The reports already on the Hub stay wrong until SmolLM2-360M fp32 is re-exported or its reports are republished.
🤖 Generated with Claude Code
Need help on this PR? Tag
@codesmith-botwith what you need. Autofix is enabled.