the phone-home register
FlashInfer
What FlashInfer sends out of your computer by default, when, to whom, and the switch that stops it, version by version, with how we know.
The answer
By its source at v0.6.18.post1, not run: when a compiled kernel FlashInfer needs is not already on disk, it downloads that kernel file from edge.urm.nvidia.com at run time; FLASHINFER_NO_DOWNLOAD=1, set before the program starts, stops the download, after which a missing kernel is an error unless you stage the kernels first with the flashinfer-cubin package; vLLM 0.30.0 brings this version in.
Builds and versions
not installed: source read at tag v0.6.18.post1On by default?yes · read in codeMeasured here?no · read in codeDate checked (UTC)2026-10-10
- Version
- 0.6.18.post1 · the source at tag v0.6.18.post1
- Where the evidence came from
- not run: the source applies to Linux with an NVIDIA card
- On by default?
- yes · read in code: when a kernel binary it needs is missing
- Measured here?
- no · read in code
- Date checked (UTC)
- 2026-10-10
- Why we say so
- flashinfer/jit/cubin_loader.py read at tag v0.6.18.post1. The source at the tag; we did not check that the released wheel matches it.
Each call out
kernel download · 0.6.18.post1
- Version · build
- 0.6.18.post1 · the source at tag v0.6.18.post1 · not run: the source applies to Linux with an NVIDIA card · not installed: source read at tag v0.6.18.post1
- Sends by default
- yes · read in code: when a kernel binary it needs is missing
- Kind
- fetches
- Trigger
- as a side effect of a local request
- When
- The code at v0.6.18.post1 downloads a kernel binary at run time, each time one it needs is not on disk; we did not run it.
- Where
- download server · edge.urm.nvidia.com
- What it sends
- a download request for the missing kernel file (request)
- The vendor's stated purpose
- “Core package that compiles/downloads kernels on first use” — README at v0.6.18.post1, the line for flashinfer-python (a list item, quoted whole), read 2026-10-10
- Also know
- vLLM 0.30.0 pins this version in its CUDA requirements; SGLang 0.5.20 pins FlashInfer 0.6.18, one release earlier. — github.com · cuda.txt, read 2026-10-10
- Grade and method
- read · flashinfer/jit/cubin_loader.py read at tag v0.6.18.post1. The source at the tag; we did not check that the released wheel matches it. · read 2026-10-10
- Receipts
- the default download address · the switch, and the download when a file is missing · the README's install options
The switch
FLASHINFER_NO_DOWNLOAD=1
- Accepted values
- 1
- What it is
- environment variable
- Where it must be set
- in the environment of the program, before it starts (see Turn them off) · turn them off
- Also turns off
- any kernel not already on disk: the library refuses to download it and stops with an error
- What turning it off costs
- it fails closed: stage the kernels first with the flashinfer-cubin package
- Proof
- found in code; not tested by us
- For the call
- kernel download · 0.6.18.post1
Listens beyond the machine, and other cloud features
Listens beyond the machine by default
- n/a · a library (0.6.18.post1 the source at tag v0.6.18.post1)
Other cloud features
None found.
What it leaves on disk
Nothing we know of.
Brings in, and installed for you by
Installed for you by
Vendor documents it? · public discussion
Vendor documents it?
- kernel download · 0.6.18.post1: partly · The README says the core package downloads kernels on first use and how to pre-install them; it does not name the host or the switch. · github.com · README.md · read 2026-10-10
Public discussion
History
No dated changes yet.
Not yet checked
Nothing named.