flashinfer-ai/flashinfer

flashinfer.nvfp4_quantize: make CuTe-DSL backend's `a_global_sf` arg a host side arg (to align CuTe-DSL backend with CUDA backend)

Open

#4,112 opened on Jul 23, 2026

View on GitHub
 (2 comments) (0 reactions) (1 assignee)Python (1,031 forks)github user discovery
good first issueop: misc

Repository metrics

Stars
 (5,756 stars)
PR merge metrics
 (Avg merge 11d 23h) (186 merged PRs in 30d)

Description

As reported: "In flashinfer.nvfp4_quantize, the type of the argument a_global_sf is inconsistent. For cuda, it is a host side value; for cute-dsl, it is a device side tensor. The function description is not clear enough. We should make it host side and passed as a function argument.

What we might be able to do relatively easily is to change the cute-dsl backend kernel to accept either (a) a device side tensor of size 1 and dtype float32 (what we support today); or (b) a single float value that trivially comes from host.  That way it would be backwards compatible for those who might be using the device-side tensor."

Contributor guide

Research direction
Examine the CuTe DSL backend kernel for flashinfer.nvfp4 quantize to understand how a global sf is currently passed. Modify the kernel to accept both a device side tensor of size 1 and a host side float value, ensuring backwards compatibility. Update the Python bindings to pass the host side float as a function argument, aligning with the CUDA backend.
Tech stack
python
Domain
backendapi
Issue type
Refactor
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
CUDAC++CuTe
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40