Repository navigation
[TEST ONLY] Consolidated Bolt stack for Microsoft benchmarking - #33991
Draft
Rui Ren (rui-ren) wants to merge 6 commits into
Draft
Rui Ren (rui-ren) wants to merge 6 commits into
Rui Ren (rui-ren) wants to merge 6 commits into
Conversation
…block-packed weights Add com.microsoft::GatherPacked (CPU and CUDA), com.microsoft::MatMulPacked (CUDA), and com.microsoft::MoEPacked (CUDA) contrib operators that consume 2D/3D block-packed quantized weight blobs (such as GGUF IQ1/IQ2/IQ3/IQ4/Q2_0/Q4/Q5/Q6_K/Q8_0) directly without offline repacking or split scale tensors. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Disallow output-type split-K reductions in MatMulPacked prefill so FP32 compute cannot silently round intermediate sums to BF16 or FP16. This addresses the Windows A10 packed-operator parity failures while retaining Tensor Core execution. Preserve the caller's cuBLAS math mode, propagate mode-update failures, and add FP16/BF16 cancellation coverage with a documented prefill numerical contract. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Validate dimensions before narrowing, support unaligned packed and activation buffers, and bound MatMulPacked prefill scratch. Check compiled BF16 capability, preserve endian-safe decoding and ggml notices, and extend CPU/CUDA regression coverage. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Resolve the generated operator documentation conflicts by retaining both MoEGroupRouter and the packed operators. Validation: CUDA 13.0 rebuild; 22 packed-operator tests, 14 focused C++ tests, 19 documentation tests; merged registration lint. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Start from current PR #33953 (2d9d48d) and consolidate the upper three PRs #33962, #33963, and #33969 into a standalone test branch. Preserve the latest packed-op validation, alignment, BF16 capability, and bounded prefill scratch fixes alongside Bolt scaling, routing, precision, and Mamba snapshots. Cover scaled prefill across N tiles. Only the new benchmark branch is changed; the original stack remains untouched. CUDA/host/schema/test translation units and formatting pass. Runtime tests are unqualified because the borrowed overlay support runtime has incompatible symbols. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 77d657ce-b25d-4ca6-9727-5653117637e0 Signed-off-by: rui-ren <ruiren1225@gmail.com>
Open
3 tasks done
This branch has not been deployed
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.
Description
TEST / BENCHMARK ONLY — do not merge.
Standalone consolidation of all four PRs in stack #33967, targeting
mainfromruiren/bolt-ms-bench:2d9d48d4897e3f26e3148f630e08624a19fab5f2ca5bf6a4c1308b4374238b37884d07edf4c3bae5a9bdbce374a3b2757e1fd11ef83ce6f83502b0d9374c7d51d93bfdd29cb0f2efbc76b88e1e812aa7The current bottom PR contains fixes and
mainupdates absent from the upper branches. This branch starts at its current head and applies the combined upper-three delta from their shared packed-op baseline. Consolidation conflicts are resolved only here, preserving dimension/byte-size checks, unaligned loads, compiled BF16 capability checks, bounded packed-weight prefill scratch, native projection precision/scaling, preselected/sigmoid MoE routing, and Mamba snapshot behavior.Includes a focused FP16/BF16 NVFP4 global-scale regression crossing the 4096-column prefill tile boundary.
Motivation and Context
Provide one independent branch and PR for Microsoft-side builds, tests, and benchmarks. The original four PRs, their head branches, base targets, draft states, and native stack membership are unchanged. This PR is not appended to the stack and does not supersede or close any source PR.
Validation
git diff --checkpasses; inherited Mamba sources and MoE routing tests are preserved exactly.