Repository navigation
Conversation
Add HRX dispatch and Loom kernels for hyperconnections, sparse attention, router reductions, and the remaining DeepSeek V4 tensor operations. Update graph caching, quantized matmul paths, tests, and operator coverage. Assisted-by: Codex
Select stream and head geometry from tensor shapes. Support gated and scaled hyperconnection inputs, identity residual mixing, and partial indexer head tiles. Include gate mode and scale in graph matching and diagnostics. Add boundary and numerical cases and refresh operator coverage. Assisted-by: Codex
jammm
marked this pull request as draft
November 1, 2026 13:38
Group attention heads, partition top-k selection and share MXFP4 expert loads across tokens. Preserve accumulation order and cover tails, routing and mask layouts. Assisted-by: Codex
Reuse float parameter formatting across scalar and hyperconnection operators. Avoid a temporary vector and redundant kernel construction in the MXFP4 path. Assisted-by: Codex
jammm
marked this pull request as ready for review
November 3, 2026 16:10
Preserve bounded graph and prepared-command caches alongside upstream weight residency and compilation diagnostics. Keep both dispatch eligibility checks and both branches' test coverage. 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Enable DeepSeek V4 Flash with DSpark on HRX/Loom. Add missing operators, sparse attention and quantized matmul paths, and generalize hyperconnection/indexer kernels. Improve decode with grouped attention, partitioned top-k selection and shared expert weight loads. Consolidate scalar parameter formatting and simplify MXFP4 dispatch setup.
Measured on gfx1151 at
64dc287eusing Unsloth DeepSeek V4 Flash 0731 UD-Q4_K_XL and Q8_0 DSpark. Throughput change:(HRX / baseline - 1) * 100. K = 1024 tokens.Matched uncached prefill; 128 greedy output tokens (127 timed), F16 KV, batch/ubatch 512, draft maximum 3,
p_min=0. Weighted warm throughput: four repetitions through 8K, two above; cold requests excluded. Vulkan uses host-preferred mmap. Small gains may overlap measurement variation.Baselines: gfx11 HIP
0a4e43a6with a separate local FP32 flash-attention accumulation fix (not included here), and upstream Vulkan9b8446b9. Tested with TheRock core/devel/libraries/device wheels10.3.0a20261025and HRX runtime7a7f1fe5plus local runtime fixes.Validation at
64dc287e: HIP and HRX pass unchanged full-logit gates at all ten contexts through 256K (NLL, KL, top-1 and high-margin predictions). The HIP 256K NLL failure is fixed. HRX passes mixed verification/rollback, 99 native operator cases, independent kernel references and access/race sanitizers. Response divergences qualify against Vulkan on matching histories (top-five tokens, log-probability regret <=0.5).Cleanup at
faddaf48: 264 canonical Loom modules and 258 full-vocabulary logit rows (512/8K contexts) match the measured build exactly. 113 backend and 146 native operator cases passed; broader standalone checks reproduce ten pre-existing SCALE/softmax support failures.AI-assisted implementation and validation.