ZacharyZcR 57706a0200 Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16) ...

* feat: add experimental CUDA backend for resident tensors

* feat: promote pinned experts to a bounded VRAM tier

* feat: preload the GPU expert tier at startup

* fix: harden CUDA backend failure handling

* feat: add deterministic multi-GPU tensor placement

* test: add deterministic CUDA benchmark fixture

* perf: make routed experts the default CUDA path

2026-07-10 07:41:09 +02:00

..

backend_cuda_test.cu

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

backend_cuda.cu

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

backend_cuda.h

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

benchmark_cuda_fixture.py

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

coli

MTP head is int8 by default (int4 head: 0-4% acceptance, unusable — measured in #8); README: honest MTP numbers + M5 Max community datapoint (#4, #5)

2026-07-09 20:34:08 +02:00

compat.h

macOS/Apple Silicon port: NEON kernels + platform shim (PR #1 by RDouglasSharp)

2026-07-06 22:01:12 +02:00

convert_fp8_to_int4.py

MTP head is int8 by default (int4 head: 0-4% acceptance, unusable — measured in #8); README: honest MTP numbers + M5 Max community datapoint (#4, #5)

2026-07-09 20:34:08 +02:00

download_glm52.py

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

eval_glm.py

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

fetch_benchmarks.py

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

gen_unicode.py

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

glm.c

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

iobench.c

macOS/Apple Silicon port: NEON kernels + platform shim (PR #1 by RDouglasSharp)

2026-07-06 22:01:12 +02:00

json.h

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

make_glm_bench_model.py

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

make_glm_oracle.py

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

Makefile

Tiered CUDA acceleration for routed experts (opt-in, CPU default untouched) + REPLAY fixture harness (#16)

2026-07-10 07:41:09 +02:00

olmoe.c

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

ref_glm.json

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

run.sh

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

setup.sh

macOS/Apple Silicon port: NEON kernels + platform shim (PR #1 by RDouglasSharp)

2026-07-06 22:01:12 +02:00

st.h

macOS/Apple Silicon port: NEON kernels + platform shim (PR #1 by RDouglasSharp)

2026-07-06 22:01:12 +02:00

supervisor.sh

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

tok_test.c

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

tok_unicode.h

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00

tok.h

colibrì: pure-C GLM-5.2 (744B MoE) engine with disk-streamed experts

2026-07-05 20:57:25 +02:00