Repository navigation
Pin gpu-kernels e85b7547: m128_w8_f2_sb prefill tiles + raster groups (27B pp +1.6..2.5%) - #490
Merged
Merged
Conversation
… (Qwen3.8-27B pp +1.6..2.5%) gpu-kernels e85b7547 (squash of #10 on top of 8f26d1f4) adds, since eb97e1dc: - #18: one header for the kqt workgroup geometry (values unchanged). - #10: m128_w8_f2_sb dense prefill tiles for Q4_K/Q5_K/Q6_K and group_m raster groups for inputs above 32 MiB of f16 (output-identical). Measured on strixhalo gfx1151, performance mode, private build (llama.cpp 5dcf6623 + hrx-system 82804197 + the add-on, the hw-gate configuration), interleaved llama-bench A/B against the 8f26d1f4 add-on, 3 passes, medians: Qwen3.8-27B UD-Q4_K_XL pp512 597.4 -> 611.7 (+2.4%), pp2048 602.0 -> 616.2 (+2.4%), pp4096 589.8 -> 604.4 (+2.5%) at ub 512; pp4096 563.2 -> 572.3 (+1.6%) at ub 2048; tg64 12.2 -> 12.2. Qwen3-4B Q5_K_M pp2048 3122.5 -> 3215.8 (+3.0%) and pp4096 2094.2 -> 2137.8 (+2.1%) at ub 2048. Correctness: test-backend-ops MUL_MAT 359/359 (x3 default, x2 tile forced, raster forced); dense-prefill-bench vs CPU bad = 0 (T 256..4097); KLD vs base 0.000000 on 27B (c512 ub512, c4096 ub2048) and ZAYA1-8B (c512, all logits through the ragged q6_K lm_head). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
added
the
hw-gate
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
gpu-kernels e85b7547 (squash of #10 on top of 8f26d1f4) adds, since eb97e1dc:
groups for inputs above 32 MiB of f16 (output-identical).
Measured on strixhalo gfx1151, performance mode, private build (llama.cpp 5dcf6623 +
hrx-system 82804197 + the add-on, the hw-gate configuration), interleaved llama-bench
A/B against the 8f26d1f4 add-on, 3 passes, medians:
Qwen3.8-27B UD-Q4_K_XL pp512 597.4 -> 611.7 (+2.4%), pp2048 602.0 -> 616.2 (+2.4%),
pp4096 589.8 -> 604.4 (+2.5%) at ub 512; pp4096 563.2 -> 572.3 (+1.6%) at ub 2048;
tg64 12.2 -> 12.2.
Qwen3-4B Q5_K_M pp2048 3122.5 -> 3215.8 (+3.0%) and pp4096 2094.2 -> 2137.8 (+2.1%)
at ub 2048.
Correctness: test-backend-ops MUL_MAT 359/359 (x3 default, x2 tile forced, raster
forced); dense-prefill-bench vs CPU bad = 0 (T 256..4097); KLD vs base 0.000000 on
27B (c512 ub512, c4096 ub2048) and ZAYA1-8B (c512, all logits through the ragged
q6_K lm_head).
Auto-merge is off on purpose: merge by hand after the hw_gate job passes.
🤖 Generated with Claude Code