FazBrowse GitHub Viewer | Trending | | Home

Tools: [Download Repo ZIP] [Original HTTPS Page]

History for benchmarks - VectorArc/avp-python · GitHub

Commits

History for

Commits on Mar 25, 2026

Complete API stability audit: all 33 findings resolved

SStas
and
claude
committed
2693fd4

Commits on Mar 23, 2026

v0.4.0 prep: remove deprecated API, rewrite README with integration examples

SStas
and
claude
committed
b70cf9e
Fix version consistency across codebase

SStas
and
claude
committed
1122891
Fix vLLM integration benchmark: point to main branch

SStas
and
claude
committed
363574d

Commits on Mar 22, 2026

Fix benchmark files for numpy hidden states, add HF sanity benchmark

SStas
and
claude
committed
341f028
Remove dead ridge/procrustes calibration code, add cross-model GGUF benchmark

SStas
and
claude
committed
91e7acc
Add OllamaConnector, production-harden LlamaCppConnector

SStas
and
claude
committed
bfb4c15

Commits on Mar 21, 2026

debug: log answer start AND end
SStas
committed
8b12028
debug: log chat template output
SStas
committed
a94852f
test: steps=0 to confirm latent steps are the issue
SStas
committed
9c1490d
Test same-prompt continuation (no different solver prompt)

SStas
and
claude
committed
e1ae511
Apply chat template in think() and generate() — critical fix

SStas
and
claude
committed
330347f
test: steps=0 KV-reuse-only (no latent projection)
SStas
committed
ae5e27d
debug: log answer length and tail
SStas
committed
a4abb5e
Increase n_ctx to 4096, max_tokens to 2048 for llama.cpp GSM8K

SStas
and
claude
committed
f281d91
Increase max_tokens to 1024 for llama.cpp GSM8K benchmark

SStas
and
claude
committed
17fd2e3
Fix CUDA: preload libcudart.so.12 via ctypes before llama_cpp import

SStas
and
claude
committed
4fcb0f0
Fix GPU: LD_LIBRARY_PATH to torch's bundled CUDA libs + debug logging

SStas
and
claude
committed
7956c09
Drop cb_eval, use embeddings=True + llama_get_embeddings_ith

SStas
and
claude
committed
e544f9a
Diagnostic: test embeddings=True context + sampling (no cb_eval)

SStas
and
claude
committed
91377b9
Add diagnostic: low-level decode+greedy WITHOUT cb_eval

SStas
and
claude
committed
9c1b543
Fix generation: skip prompt decode for same-prompt continuation,

SStas
and
claude
committed
f721493
test: steps=0 same-prompt continuation diagnostic
SStas
committed
a38513e
Add llama.cpp GSM8K benchmark with GPU/CPU auto-fallback

SStas
and
claude
committed
c85bd60
Switch llama.cpp benchmark to CPU-only (CUDA wheel linking issues)

SStas
and
claude
committed
13c9cc2
Simplify CUDA setup: let torch provide CUDA runtime, then install CUDA wheel

SStas
and
claude
committed
8f83ea3
Add libgomp1 for OpenMP support in llama-cpp-python CUDA wheel

SStas
and
claude
committed
9d8748b
Use CUDA runtime base image for libcudart.so

SStas
and
claude
committed
35cd236
Use pre-built CUDA wheel for llama-cpp-python (no compilation needed)

SStas
and
claude
committed
843e0b7
Fix CUDA build: override CC/CXX to gcc (clang not in CUDA image)

SStas
and
claude
committed
8cf7e46
Use CUDA devel base image for llama-cpp-python GPU build

SStas
and
claude
committed
fcbb484
Build llama-cpp-python with CUDA support on Modal

SStas
and
claude
committed
6b6db6e
Add GGUF weight extraction diagnostic to llama.cpp benchmark

SStas
and
claude
committed
1a70854
Add gguf to Modal benchmark image for embed weight extraction

SStas
and
claude
committed
38c3e13
Fix remaining hidden_state → last_hidden_state in benchmark

SStas
and
claude
committed
b59b8b4

Back | FazBrowse Home | New Git URL