Repository navigation
Vendor ggml b11913; report weight types in /info - #3
Merged
Merged
Conversation
Update ggml from llama.cpp b8833 (2026-04) to b11913 (2026-10). The local overlay (GTE ops, repack tweaks, llamafile fallback fix, WebGPU GTE kernels) is re-applied with a 3-way merge against pristine b8833. The WebGPU adapter checks now follow upstream, which rejects adapters without shader-f16 at registration. CPU output is unchanged (cos 1.000000 against b8833 on bge-base-zh, MiniLM, text2vec, Snowflake fp16/q4_k_mlp_q8_attn/q4_0 and bge-m3 fp16/q8_0/mixed). Speed on a Xeon E5-2673 v3 (AVX2, 12 threads): bge-m3 q8_0 +7%, q4_k MLP mix +9%; Apple M4 CPU unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mixed quantizations share the file-level ftype (bge-m3 q8_0 and q4_k MLP + q8_0 attention both report "q8_0"), so clients that key vector spaces on /info could not tell them apart. /info now also has weight_types, the matrix weight types with counts (e.g. "q4_K:49,q8_0:97"), logged at load as well. README: the two-pass EMBEDDINGS_CPP_SKIP_QUANT_PATTERNS recipe for bge-m3 (q4_k MLP and vocabulary, q8_0 attention: q8_0 retrieval quality, ~40% faster on AVX2, 384 MB). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The b11913 merge kept the local early returns in create_webgpu_device, which upstream made void. Return bool again so ggml_backend_webgpu_reg_get_device can report no device rather than abort. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The second argument of page.waitForFunction is the page function's argument, so the 120 s timeout was ignored and the default 30 s applied; loading the model and encoding on a CI runner takes longer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Headless Chromium now rejects the ~190 MB IndexedDB write of the Snowflake model, and the encode page reported prepare-error, so Browser E2E has timed out since October. Skip the cache (logging why) instead of failing; Cache API writes likewise. 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
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.
/infogainsweight_types(e.g."q4_K:49,q8_0:97") so mixed quantizations, which share the file-levelftype, can be told apart.Checks so far:
🤖 Generated with Claude Code