millaguie

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-11 08:51:30 +02:00

e9616d415f cuda : gate DT3 MMQ by batch size, cuBLAS keeps the large-batch prefill

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-11 08:33:11 +02:00

981f439ff3 tests : judge the batched DT3 path strictly when it is MMQ

f46e8072d8 cuda : add MMQ kernel for DT3

Compare 2 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-11 03:31:48 +02:00

9622c56b0e tests : accept either accumulator precision in the DT3 GEMM bit-identity gate

ad6dd747d4 vulkan : fp32 accumulators for the DT3 dequant matmul fallback

8ba4db150f vulkan : add DT3 dequant, get_rows and scalar mul_mat_vec

Compare 3 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 23:33:24 +02:00

c01c26b56e tests : skip test-dt3-gpu on backends that do not implement DT3

10e1fe3d3c cuda : decode DT3 bytes once in the MMVQ vec_dot

0f33afbe56 tests : declare the generic DT3 vec_dot weak in the parity test

4e109bc7e6 tests : check the arch DT3 vec_dot is bit-identical to the generic

bf4eca0eb6 ggml : add AVX-512 DT3 vec_dot

Compare 29 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 23:09:24 +02:00

470840e01e tests : skip test-dt3-gpu on backends that do not implement DT3

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 21:52:04 +02:00

3c8fd43e7f cuda : decode DT3 bytes once in the MMVQ vec_dot

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 18:52:13 +02:00

fbc881b9a0 tests : declare the generic DT3 vec_dot weak in the parity test

3debebfbb7 tests : check the arch DT3 vec_dot is bit-identical to the generic

b110945afc ggml : add AVX-512 DT3 vec_dot

Compare 3 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 15:20:36 +02:00

51c6b67e8b tests : gate the DT3 GEMM fallback by bit-identity with an F16 GEMM

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 15:17:50 +02:00

b8ad6c3844 tests : judge DT3 mul_mat on norm error, add fp16 reference and controls

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 15:01:43 +02:00

ca2cb04393 tests : add DT3 GPU vs CPU parity test

bc3e949e71 cuda : add DT3 MMVQ kernel

4dfb210207 cuda : add DT3 dequantization

Compare 3 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 13:46:22 +02:00

0ecb2a521e tests : check DT3 byte positions against hand-computed literals

f39d565d1a llama : warn when quantizing to DT3

d8feee2542 ggml : harden DT3 validation and reference quantizer

Compare 3 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 13:21:53 +02:00

bbc407139b tests : add bit-level DT3 tests and Rust parity driver

b3323108b5 gguf-py : add DT3

e1fd6e0a13 llama : register the DT3 file type

d4343d0c5c ggml-cpu : add DT3 generic vec_dot and type traits

948f51d274 ggml : add DT3 reference quantization and dequantization

Compare 5 commits »

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 12:58:51 +02:00

4d9a6f55b6 ggml: add block_dt3, the dual-plane ternary block

millaguie pushed to dt3 at millaguie/llama.cpp

2026-08-10 12:36:35 +02:00

millaguie created branch dt3 in millaguie/llama.cpp

2026-08-10 12:36:35 +02:00

millaguie pushed to master at millaguie/llama.cpp

2026-08-10 12:36:28 +02:00

millaguie created branch master in millaguie/llama.cpp

2026-08-10 12:36:28 +02:00

millaguie created repository millaguie/llama.cpp

2026-08-10 12:35:57 +02:00

millaguie deleted branch renovate/node-24.x from millaguie/sternboard

2026-07-08 00:43:55 +02:00

millaguie pushed to main at millaguie/sternboard

2026-07-08 00:43:55 +02:00

dc2d995873 Merge PR #17: Node 24

e65f90ee80 build(deps): Node.js 22 -> 24 (Dockerfile + CI). Closes #9

Compare 2 commits »