server: skip device enumeration in router mode to avoid creating CUDA…#23137
Merged
allozaur merged 1 commit intoMay 16, 2026
Merged
Conversation
Contributor
|
Can confirm that it fixes the problem. 👍 |
Contributor
Author
|
Since log handling has already been improved, adding an explicit commented argument should, I think, prevent any regression. |
ngxson
approved these changes
May 16, 2026
allozaur
approved these changes
May 16, 2026
kgrama
pushed a commit
to kgrama/llama.cpp
that referenced
this pull request
May 19, 2026
xxmustafacooTR
pushed a commit
to xxPlayground/llama-cpp-turboquant
that referenced
this pull request
May 19, 2026
rsenthilkumar6
pushed a commit
to rsenthilkumar6/llama.cpp
that referenced
this pull request
May 19, 2026
ArberSephirotheca
pushed a commit
to ArberSephirotheca/llama.cpp
that referenced
this pull request
May 19, 2026
baramofme
pushed a commit
to baramofme/llama-cpp-turboquant
that referenced
this pull request
May 23, 2026
srossitto79
pushed a commit
to srossitto79/llama.cpp
that referenced
this pull request
May 23, 2026
winstonma
pushed a commit
to winstonma/llama.cpp
that referenced
this pull request
May 27, 2026
fewtarius
pushed a commit
to fewtarius/llama.cpp
that referenced
this pull request
May 30, 2026
Jcfunk
added a commit
to Jcfunk/llama.cpp
that referenced
this pull request
Jun 2, 2026
* turboquant/HEAD: (82 commits) docs(readme): credit Google's original TurboQuant + explain the '+' docs(readme): fix turbo ladder ordering + cite K-compression paper docs(readme): reorder KV configs as a ladder + 'start light' guidance docs(readme): add Chronara to deployments + AtomicChat link docs: restructure README — professional layout, deployments, paper links docs: tighten README — add turbo2, missing features, paper links docs: keep upstream README, prepend fork-specific summary docs: replace upstream README with fork-specific summary fix(xxd.cmake): handle missing input file (not just empty) fix(ci): 4 cross-vendor -Werror failures + defensive xxd.cmake cmake : fix LLAMA_BUILD_UI logic (ggml-org#23190) fix(ggml-cuda): HIP nodiscard + MUSA cudaMemcpyToSymbol alias fix(turbo-quant): add forward declaration for turbo_cpu_fwht_inverse fix(metal): set ne12/ne13/r2/r3 function constants in mul_mm_tq_rotated pipeline webui: support video files as input (ggml-org#22830) server: (router) alloc tmp buffer on heap (ggml-org#23159) server: skip device enumeration in router mode to avoid creating CUDA primary context (ggml-org#23137) vulkan: removed duplicate #include <memory> in headers (ggml-org#23144) ui: Add request timeout for MCP tool calls (ggml-org#23138) sync : ggml ...
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
4 participants
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
In router mode (no model loaded), the parent llama-server eats 500MiB of VRAM per CUDA device. (The pre-allocated memory amount is related to GPU SMs number. The GPU with more SMs requires a larger memory)
Fix this
Additional information
Regression from #23021, which moved the device info log into common_params_print_info and wired it in main before the router mode gets detected.
The culprit is ggml_backend_dev_memory inside the loop: on CUDA it lands in cudaSetDevice + cudaMemGetInfo, which materialize the primary context on each device. The router parent never touches the GPU, so the context is pure waste and the PID shows up in nvidia-smi. Invisible on Metal because of unified memory, which is why it slipped through.
Fix adds an optional print_devices flag (default true, no other caller touched). Router passes false. Build info, verbosity and system_info still log normally, only the device loop is skipped.
Same class of bug previously fixed in #20595 (closes #20582). Closes #23120.
router log
child log
Requirements