Donato Capitella
227fa190aa
perf(rocm): accelerate gfx1151 DSpark verification
2026-08-29 21:51:45 +01:00
Donato Capitella
307dc093d9
perf(rocm): specialize tiny Q8 verifier batches
2026-08-29 21:51:45 +01:00
Donato Capitella
a4e015e004
perf(rocm): accelerate tiny-batch DSpark verification
2026-08-29 21:51:45 +01:00
Donato Capitella
ab3346a695
bench: finalize gfx1151 prefill result
2026-08-29 21:51:45 +01:00
Donato Capitella
1ef0680b79
perf(rocm): default to tuned gfx1151 prefill paths
2026-08-29 21:51:45 +01:00
Donato Capitella
49a30025d6
perf(rocm): reduce gfx1151 prefill staging overhead
2026-08-29 21:51:45 +01:00
Donato Capitella
b7f2f2d37f
bench: close gfx1151 attention key tile sweep
2026-08-29 21:51:45 +01:00
Donato Capitella
86128aae00
bench: record 290.79 tok/s gfx1151 prefill lead
2026-08-29 21:51:45 +01:00
Donato Capitella
4e23cf1bcc
rocm: use 80-key WMMA attention tiles on gfx1151
2026-08-29 21:51:45 +01:00
Donato Capitella
becf9f96d8
bench: record large F16 gfx1151 win
2026-08-29 21:51:45 +01:00
Donato Capitella
8c1629c95c
perf(rocm): accelerate large F16 prefill projection
2026-08-29 21:51:45 +01:00
Donato Capitella
491e7e3d3f
perf(rocm): remove rejected paired IQ2 kernel
2026-08-29 21:51:45 +01:00
Donato Capitella
9fd5e399e9
bench: profile paired gfx1151 IQ2 kernel
2026-08-29 21:51:45 +01:00
Donato Capitella
c40e8a8e09
perf(rocm): pair IQ2 gate and up tiles
2026-08-29 21:51:45 +01:00
Donato Capitella
11e1eb31b0
bench: record compact gfx1151 IQ2 launch
2026-08-29 21:51:45 +01:00
Donato Capitella
4bf7c47b71
perf(rocm): compact IQ2 MoE tile launches
2026-08-29 21:51:45 +01:00
Donato Capitella
540ccbafa7
bench: record 277 tok/s gfx1151 lead
2026-08-29 21:51:45 +01:00
Donato Capitella
9ab922f50e
perf(rocm): tighten IQ2 MoE launch bound
2026-08-29 21:51:45 +01:00
Donato Capitella
45dbb31505
docs: record gfx1151 small-M F16 results
2026-08-29 21:51:45 +01:00
Donato Capitella
6bdc68bccf
perf(rocm): specialize small-M F16 prefill GEMMs
2026-08-29 21:51:45 +01:00
kyuz0
b0d80b13c8
bench: record 263 tok/s gfx1151 lead
2026-08-29 21:51:45 +01:00
kyuz0
e4cd58705b
rocm: specialize tiny-m f16 prefill gemm
2026-08-29 21:51:45 +01:00
kyuz0
d8d5fa679a
bench: record 256 tok/s gfx1151 lead
2026-08-29 21:51:45 +01:00
kyuz0
4d89fe9e84
rocm: vectorize indexed attention kv staging
2026-08-29 21:51:45 +01:00
kyuz0
df1269414d
bench: close unsafe f16 kv mirror experiment
2026-08-29 21:51:45 +01:00
kyuz0
64b22d74f1
bench: map remaining gfx1151 prefill pools
2026-08-29 21:51:45 +01:00
kyuz0
300c5814b7
bench: record indexed attention tile rejection
2026-08-29 21:51:45 +01:00
kyuz0
68e0ca1e44
bench: record 248 tok/s prefill lead
2026-08-29 21:51:45 +01:00
kyuz0
e43738a0ab
rocm: specialize attention output B for gfx1151
2026-08-29 21:51:45 +01:00
kyuz0
d104a31b14
bench: map attention output GEMM
2026-08-29 21:51:44 +01:00
kyuz0
c2b67652e9
bench: record Q2 n4 rejection
2026-08-29 21:51:44 +01:00
kyuz0
f6022ee42d
bench: record IQ2 sign-mask optimization
2026-08-29 21:51:44 +01:00
kyuz0
0610a6577e
rocm: use cached IQ2 sign masks in MMQ
2026-08-29 21:51:44 +01:00
kyuz0
dd356c3fad
bench: record gfx1151 prefill optimization results
2026-08-29 21:51:44 +01:00
kyuz0
7470dd60a7
build: link MMQ objects in ROCm regression target
2026-08-29 21:51:44 +01:00
kyuz0
c53b2ef6fa
rocm: accelerate sparse attention with WMMA
2026-08-29 21:51:44 +01:00
kyuz0
b93fb08937
rocm: tile Q8 prefill WMMA by output shape
2026-08-29 21:51:44 +01:00
kyuz0
9933e01f50
rocm: wire tuned MMQ IQ2 prefill experiment
2026-08-29 21:51:44 +01:00
kyuz0
61e9b991c3
rocm: add gfx1151 MMQ kernel lab
2026-08-29 21:51:44 +01:00
kyuz0
e6e7e6a5d7
rocm: reuse LDS output tile in Q2_K MoE down kernel
2026-08-29 21:51:44 +01:00
esvhd
8db89fe083
download: add GLM 5.3 Flash models ( #888 )
...
Adds Hugging Face download targets for the Q2, Q4_K, and native FP8 artifacts.
Co-authored-by: esvhd <7330385+esvhd@users.noreply.github.com >
2026-08-28 23:25:07 +02:00
antirez
c1d4597a80
qa: update DGX Spark host addresses
2026-08-23 15:26:19 +02:00
antirez
c35cf38ba9
models: download the DeepSeek V4 Pro 0813 quant
...
Point pro-q2-imatrix at the published dated GGUF and update the README examples and native PRO conversion description.
2026-08-23 15:05:13 +02:00
antirez
3c63c06c5a
qa: add DeepSeek V4 Pro 0813 oracle
...
Keep the preview oracle intact, add the 100-case official 0813 fixture, and record the source, conversion, safety, quality, hash, and publication results.
2026-08-23 15:05:13 +02:00
antirez
ed5c1a92c6
gguf: harden DeepSeek PRO quantization
...
Load compression ratios from the source checkpoint instead of retaining stale template metadata. Stop Q2_K/Q4_K refinement when an accepted fit collapses the search range.
2026-08-23 15:05:13 +02:00
antirez
abd4ed3885
cli: expose exact MTP sampling mode
2026-08-23 15:05:13 +02:00
antirez
8d683d6616
dspark: add opportunistic temperature sampling
2026-08-23 15:05:13 +02:00
antirez
769a8ba42d
dspark: support exact stochastic speculative decoding
2026-08-23 15:05:13 +02:00
Salvatore Sanfilippo
ecdea42ee7
Merge pull request #656 from jyatesdotdev/rocm-mxfp4-pr
...
Add ROCm MXFP4 routed expert kernels
2026-08-23 15:03:33 +02:00
Jonathan Yates
babca3d0fe
Makefile: restore Darwin link guard for ds4_test/ds4_agent_test
2026-08-20 21:23:53 -07:00
Donato Capitella
kyuz0
esvhd
antirez
Jonathan Yates