Repository navigation
Conversation
…ce by default on Alchemist (bitwise) At 128 registers the step kernel spilled ~7.6 KB per SIMD32 thread on the A770 (unitrace). STRATA_GDN_LARGE_GRF (Alchemist default) launches the recurrence kernels in large-GRF mode; STRATA_GDN_SPLIT (S25/S26, bitwise) becomes the Alchemist default. gdn_bench (new: 48 layers' states from VRAM), 4 / 6 tokens: 103 / 114 -> 39.5 / 44.9 us; gdn_parity 0 failures; engine 8K GPU wait 39.5 / 38.8 -> 37.6 / 37.9 (split) and 35.9-36.3 -> 33.8-34.0 ms/round (large GRF). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…launches (STRATA_GDN_COMMIT_BATCH; bitwise) capture_commit replayed the accepted tokens into each GDN layer's state with two launches a layer (gdn_conv_commit and the split state replay): 36 layers -> 72 launches at the end of every verify window, on the critical path into the next round's drafting. gdn_commit_replay_layers runs the same two kernels over all layers at once (the layer is the outermost work-group index, selecting each layer's state, history and inputs), so every thread computes what it did. It declines (the per-layer loop runs) where the split replay is not the path. Default on on Alchemist, where it is measured; STRATA_GDN_COMMIT_BATCH=1 forces it on, =0 off. Needs the split replay (Niko1221#2586). gdn_bench GDN_CBATCH=1 (new): 36 layers, n_keep 1..6, every float of every layer's state and history compared per layer vs batched: 0 differ; 1.96 -> 0.83 ms. Engine, A770, Coder IQ1_M, 8K code prompt, 3 pairs: MTP drafting 7.51 / 7.31 / 7.44 -> 6.58 / 6.44 / 6.50 ms a round (-0.91), decode 35.68 / 39.41 / 39.58 -> 35.58 / 40.04 / 41.63 tok/s, reference tokens (one known low-hit variant). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 19, 2026
Owner
|
Thanks for the pull request. We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected. To reopen it takes about two minutes:
The rules are in CONTRIBUTING.md. |
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.
sycl/A770: commit all GDN layers' conv history + state replay in two launches (STRATA_GDN_COMMIT_BATCH; bitwise)
capture_commit replayed the accepted tokens into each GDN layer's state with two launches a layer (gdn_conv_commit and
the split state replay): 36 layers -> 72 launches at the end of every verify window, on the critical path into the
next round's drafting. gdn_commit_replay_layers runs the same two kernels over all layers at once (the layer is the
outermost work-group index, selecting each layer's state, history and inputs), so every thread computes what it did.
It declines (the per-layer loop runs) where the split replay is not the path. Default on on Alchemist, where it is
measured; STRATA_GDN_COMMIT_BATCH=1 forces it on, =0 off. Needs the split replay (#2586).
gdn_bench GDN_CBATCH=1 (new): 36 layers, n_keep 1..6, every float of every layer's state and history compared per
layer vs batched: 0 differ; 1.96 -> 0.83 ms. Engine, A770, Coder IQ1_M, 8K code prompt, 3 pairs: MTP drafting
7.51 / 7.31 / 7.44 -> 6.58 / 6.44 / 6.50 ms a round (-0.91), decode 35.68 / 39.41 / 39.58 -> 35.58 / 40.04 / 41.63 tok/s,
reference tokens (one known low-hit variant).
Stacked on #2586 (the split GDN replay, which this batches); the diff to review is the top commit.
🤖 Generated with Claude Code