Skip to content

Navigation Menu

Appearance settings

View all features
- BY COMPANY SIZE
  Enterprises
  Small and medium teams
  Startups
  Nonprofits
- BY USE CASE
  App Modernization
  DevSecOps
  DevOps
  CI/CD
  View all use cases
- BY INDUSTRY
  Healthcare
  Financial services
  Manufacturing
  Government
  View all industries
View all solutions
- EXPLORE BY TOPIC
  AI
  Software Development
  DevOps
  Security
  View all topics
- EXPLORE BY TYPE
  Customer stories
  Events & webinars
  Ebooks & reports
  Business insights
  GitHub Skills
- SUPPORT & SERVICES
  Documentation
  Customer support
  Community forum
  Trust center
  Partners
View all resources
- COMMUNITY
  GitHub SponsorsFund open source developers
- PROGRAMS
  Security Lab
  Maintainer Community
  Accelerator
  GitHub Stars
  Archive Program
- REPOSITORIES
  Topics
  Trending
  Collections
- ENTERPRISE SOLUTIONS
  Enterprise platformAI-powered developer platform
- AVAILABLE ADD-ONS
  GitHub Advanced SecurityEnterprise-grade security features
  Copilot for BusinessEnterprise-grade AI features
  Premium SupportEnterprise-grade 24/7 support
Pricing

Search code, repositories, users, issues, pull requests...

Search

Clear

Search syntax tips

Provide feedback

We read every piece of feedback, and take your input very seriously.

Include my email address so I can be contacted

Saved searches

Use saved searches to filter your results more quickly

Name

Query

To see all available qualifiers, see our documentation.

Appearance settings

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

groxaxo / GPTQ-Pro Public

forked from ModelCloud/GPTQModel

Notifications You must be signed in to change notification settings
Fork 2
Star 13

Code
Pull requests
Actions
Projects
Security and quality
Insights

Additional navigation options

Code
Pull requests
Actions
Projects
Security and quality
Insights

Optimize GPTQ-Pro kernels for Ampere decode and prefill - #11

#11

Merged

groxaxo merged 41 commits into

maingroxaxo/GPTQ-Pro:mainfrom

perf/ampere-kernel-pipelinegroxaxo/GPTQ-Pro:perf/ampere-kernel-pipelineCopy head branch name to clipboard

Jul 11, 2026

Conversation Commits41 (41)Checks Files changed

Merged

Optimize GPTQ-Pro kernels for Ampere decode and prefill#11
groxaxo merged 41 commits into
mainfrom
perf/ampere-kernel-pipeline

Commits

Commits on Jul 11, 2026

perf(cuda): add Ampere pipeline and decode primitives
groxaxo
committed
perf(cuda): add async MMA pipeline and small-M GEMV
groxaxo
committed
feat(cuda): expose kernel dispatch modes to PyTorch
groxaxo
committed
perf(runtime): reduce GPTQ-Pro packing memory and expose dispatch override
groxaxo
committed
fix(runtime): handle channelwise groups and document optimized dispatch
groxaxo
committed
fix(cuda): preserve fallback tails and FP16 dequant semantics
groxaxo
committed
fix(cuda): satisfy kernel-mode parser control flow
groxaxo
committed
test(cuda): validate every GPTQ-Pro dispatch path
groxaxo
committed
ci(cuda): compile Ampere kernel and standalone validator
groxaxo
committed
feat(bench): add raw GPTQ-Pro kernel benchmark and parity checks
groxaxo
committed
test(kernel): lock Ampere dispatch, packing, and source contracts
groxaxo
committed
docs(readme): document optimized Ampere dispatch and benchmarking
groxaxo
committed
docs(roadmap): record implemented Ampere pipeline and remaining gates
groxaxo
committed
ci(kernel): add targeted Python style and syntax gate
groxaxo
committed
ci(temp): capture formatted kernel Python sources
groxaxo
committed
ci(temp): expose format artifact on pull request runs
groxaxo
committed
style(kernel): apply Ruff formatting to runtime helper
groxaxo
committed
style(kernel): apply Ruff formatting to quantized linear runtime
groxaxo
committed
style(kernel): apply Ruff formatting to raw benchmark
groxaxo
committed
style(kernel): apply Ruff formatting to Ampere contracts
groxaxo
committed
ci(kernel): compile the real PyTorch CUDA extension binding
groxaxo
committed
test(kernel): isolate runtime helper tests from package-wide imports
groxaxo
committed
fix(ci): build PyTorch extension inside isolated environment
groxaxo
committed
fix(bench): disable TF32 in numerical reference
groxaxo
committed
perf(cuda): consume native GPTQ int32 words directly
groxaxo
committed
perf(cuda): run all kernels directly from native qweight
groxaxo
committed
perf(cuda): bind native int32 qweight without repacking
groxaxo
committed
perf(runtime): pass native qweight directly to CUDA
groxaxo
committed
perf(runtime): remove duplicate packed-weight buffer
groxaxo
committed
test(cuda): validate native qweight dispatch paths
groxaxo
committed
perf(bench): benchmark native qweight without duplicate packing
groxaxo
committed
test(kernel): lock native-qweight runtime contract
groxaxo
committed
docs(readme): document native-qweight Ampere runtime
groxaxo
committed
docs(roadmap): record native-qweight runtime and compile gates
groxaxo
committed
style(kernel): apply exact Ruff format after native-qweight pivot
groxaxo
committed
style(kernel): apply exact Ruff format to native-qweight tests
groxaxo
committed
chore(ci): remove temporary kernel format diagnostics
groxaxo
committed
fix(runtime): normalize native qweight and scale contiguity once
groxaxo
committed
chore(cuda): include fixed-width integer declarations explicitly
groxaxo
committed
fix(cuda): version the extension ABI after native-qweight changes
groxaxo
committed
fix(ci): isolate kernel contracts from repository conftest
groxaxo
committed

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms
Privacy
Security
Status
Community
Docs
Contact

You can’t perform that action at this time.