Repository navigation
rdna35 prefill optimizations other than mmq - #129
Draft
liangliangchang wants to merge 14 commits into
Draft
liangliangchang wants to merge 14 commits into
liangliangchang wants to merge 14 commits into
Conversation
Assisted-by: Codex
Assisted-by: Codex
Assisted-by: Codex
Assisted-by: Codex
Assisted-by: Codex
Batch Qwen vision windows and tune guarded RDNA3.5 attention paths. Fuse clamp conversion, normalization and activation packing, and residual scaling while preserving existing dispatch and rounding. Add backend coverage and document the matched VLM performance results. Assisted-by: Codex
liangliangchang
changed the title
Zero-pad compatible D72 self-attention to the existing D80 WMMA kernel, retain the original attention scale, and extract D72 output. Guard small workloads, unsupported layouts and attention features. Add backend coverage for dispatch boundaries and fallback paths. On gfx1151, Gemma-3-4B median HTTP TTFT improves from 1141.47 ms to 769.71 ms with context checkpoints enabled (32), using 20 timed requests per configuration in ABBA order. D72 and D64/D80 backend suites pass 909 and 224 cases; 32 generated token IDs match baseline. Floating-point log probabilities differ with the WMMA arithmetic. Assisted-by: Codex
Extend shared GLU producer kernels to SwiGLU and padded Q8 rows. Use 128-thread blocks for eligible D128 Q/K norm-RoPE prefill. Preserve existing dispatch, FP32 rounding and fallback paths. On gfx1151 with checkpoints enabled, 3968-token HTTP TTFT falls 4.76% for Qwen2.5-3B and 3.85% for Qwen3-4B. Producer, graph and generation checks pass with the recorded matching outputs. Assisted-by: Codex
Document every branch optimization commit, structural eligibility, measurement baseline and performance limits. Include the SwiGLU results and distinguish isolated gains from combined experiments. Assisted-by: Codex
jimw567
self-requested a review
October 14, 2026 22:21
Share gate/up input conversion and retain FP16 GEMM intermediates. Fuse conversion, bias and gated activation with FP32 arithmetic and existing rounding, dispatch and fallback behavior. Add backend coverage and document measured TTFT gains and scratch-memory costs. Qwen2.5-VL-7B saves 5.1-5.7% TTFT; Gemma-3-4B saves 3.2% with checkpoints enabled. All 104 backend cases and direct numerical checks pass, with identical tested responses and returned probabilities. Assisted-by: Codex
Reuse conversion/bias, GLU, gated FFN and clamp fusions for BF16. Preserve FP32 arithmetic, rounding and existing dispatch guards, including fallback for mixed projection types. Add BF16 backend coverage and document numerical validation, paired TTFT gains and the completed 14-row VLM comparison. Assisted-by: Codex
Hoist window K/V casts, write attention into the final output, and extend normalization and residual producer fusions. Document Qwen2.5-VL-7B timing, checkpoint-on validation and numerical differences. Assisted-by: OpenAI Codex
Extend affine LayerNorm, plain GELU and clamped gated FFN fusions, and combine compatible two-axis norm-RoPE operations. Retain FP32 producer rounding, clamp limits, dispatch guards and checkpoint behavior. Document model coverage, paired TTFT results, numerical checks and rejected variants. Assisted-by: Codex
Skip partial-tile handling for complete dense D80 KV tiles, vectorize aligned permuted F32-to-F16 copies, and use 128-thread affine LayerNorm for 1152-wide vision rows. Keep final-tile checks and guarded fallbacks. Document Gemma-3-12B TTFT of 857.01 to 818.46 ms with checkpoints and overlap enabled, host-load limits, and LayerNorm rounding differences. Assisted-by: Codex
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.
Overview
Additional information
Requirements