Repository navigation

feat(runtime): mmap-backed zero-copy ASR weight loading - #47

Merged

pskrunner14 merged 10 commits into

NVIDIA:mainfrom

ryanleary:asr-mmap-zerocopy

Oct 20, 2026

Merged

feat(runtime): mmap-backed zero-copy ASR weight loading#47
pskrunner14 merged 10 commits into
NVIDIA:mainfrom
ryanleary:asr-mmap-zerocopy

Commits on Oct 20, 2026
- feat(runtime): mmap-backed zero-copy ASR weight loading
  
  authored and
  pskrunner14
  committed
  a81549b
- fix(runtime): route RelPositionalEncoding's GGUF-path .pe write through the zero-copy-aware bind helper
  
  authored and
  pskrunner14
  committed
  78d96b4
- test(asr): verify GGUFLoader's mmap path against the buffered-read path
  
  authored and
  pskrunner14
  committed
  1720c95
- style: trim comments in the mmap zero-copy change down to 2 lines
  
  authored and
  pskrunner14
  committed
  71f1873
- simplify(test): drop GGUFLoader::tensor_names(), enumerate via gguf_get_tensor_name directly
  
  authored and
  pskrunner14
  committed
  bdfb13a
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  
  pre-commit-ci[bot]
  authored and
  pskrunner14
  committed
  4b09593
- fix: address CodeRabbit review on #47
  
  authored and
  pskrunner14
  committed
  3bf8825
- fix(runtime): make CPU weights eligible for zero-copy mmap binding
  
  authored and
  pskrunner14
  committed
  decc3ca
- fix(runtime): harden GGUF mmap loading against bad files and mmap failure
  
  authored and
  pskrunner14
  committed
  6970b22
- build: use NEMO_SPEECH_LLAMA_CPP_DIR for llama-mmap, register mmap test
  
  authored and
  pskrunner14
  committed
  5640db2