Repository navigation
[CPU][S390x] Vectorise topk topp for s390x - #1
Open
R3hankhan123 wants to merge 4 commits into
Open
R3hankhan123 wants to merge 4 commits into
R3hankhan123 wants to merge 4 commits into
Conversation
) (cherry picked from commit 9ffd571) Signed-off-by: Tianmu Li <tianmu.li@intel.com>
(cherry picked from commit 267cbdf) Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Remove the AVX512 and AVX2 sampling paths introduced by the base implementation while retaining the AArch64 implementations and scalar fallbacks. Remove the added AVX2 fast-exp macro and -mfma flag, apply repository formatting, and run the CPU top-k/top-p tests in the Arm CI kernel shard. Assisted-by: OpenAI Signed-off-by: Michael Yang <michael.yang@arm.com>
R3hankhan123
changed the title
Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com>
R3hankhan123
force-pushed
the
pr-62440-s390x-topk
branch
from
November 2, 2026 11:02
25c3d61 to
73f15af
Compare
R3hankhan123
added a commit
to R3hankhan123/spyre-inference
that referenced
this pull request
Nov 3, 2026
…d DFlash Split speculative decoding into a shared verification layer and per-method plug-ins, so each method stops adding branches to the platform and runner. - spec_decode/base.py: SpyreSpecMethod, whose hooks default to upstream's behaviour; registry.py resolves the config to a method and rejects any other with the supported list. - Methods: n-gram and suffix (model-free), Gemma-4 MTP (assistant checkpoint reading the target's KV cache) and DFlash/XPress, moved behind the hooks at its existing call points. - Shared verification (from torch-spyre#1446): verify compile sizes and query buckets, scheduler draft padding, and the batched decode kernel at the verify width. A drafter that attends through the target's builders (MTP) also records the kernel at its own width. - Rejection sampling on the native C++ kernels; recovered tokens are drawn with fused_gumbel_argmax over log(residual). - Gumbel noise: vLLM's counter-based draw from csrc/cpu/sampling/common.h, vendored as csrc/cpu/sampling_noise.hpp and keyed on (seed, position, token), inside Spyre's exactly pruned argmax loop. - Top-k / top-p: vLLM's CPU kernels (micyan01/vllm#1) vendored as csrc/cpu/topk_topp_sampler.cpp, without the AArch64 paths, with an AVX2 x86 path, and keeping ties at the k-th value as vLLM's PyTorch top-k does. Co-authored-by: David Holtz <56723830+dmholtz@users.noreply.github.com> Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com>
micyan01
force-pushed
the
arm-neon-topk-topp
branch
from
November 3, 2026 08:57
5fe2b34 to
ca0b7a6
Compare
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
Claims
Validation
Details
Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.
BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)