Morph MORPH
SIGN IN SIGN UP
dhruvhead / exllamav3 UNCLAIMED
0 0 0 Python

COMMITS

master
July 27, 2026
T
Bump to v1.2.1
turboderp committed
787d158
T
Util: Add tiktoken converter
turboderp committed
dfcf267
T
CPU offload: Don't sleep threads under Windows (use WakeByAddress instead), add handoff profiler, apply GPU pointer alias to pinned buffer
turboderp committed
268029a
R
Apply /DNOMINMAX fix to JIT extension load (was only applied to setup.py)
remichu committed
1d13540
July 26, 2026
T
MGEMM: Change long* to int64_t* (different type in MSVC)
turboderp committed
3ee9b0c
T
Generator: Account for vocab merges that straddle character boundaries. Fixes #255
turboderp committed
0f3b0ab
T
convert.py: Split reference forward pass across active devices, batch-convert small tensors on single device to saturate GPU on each tile row, updated scale search with outlier-aware sample selection
turboderp committed
da9688a
T
qbench.py: Add feedback
turboderp committed
07e177e
T
Remove defunct test script
turboderp committed
02f34e4
T
qbench.py: Fix tensor name regex matching
turboderp committed
61fd2f0
T
chat.py: Use platform-aware path for history file
turboderp committed
a07b95a
July 25, 2026
T
qbench.py: Small tweaks
turboderp committed
4f95e45
T
qbench.py: Fix VRAM/bpw estimate for split GGUF files
turboderp committed
1f18f21
T
qbench.py: Increase histogram sidecars to fp32 precision
turboderp committed
c9f77e2
T
qbench.py: Add self-generated trace measurement
turboderp committed
c4a7c10
T
qbench.py: Support MXFP4, option to prepend chat response when formatting
turboderp committed
aede042
T
CPU MoE: Attempt to fix Windows build
turboderp committed
ebbffea
T
SWA: Zero initial state allocation
turboderp committed
1a0ee3a
T
Bump to v1.2.0 (pending Windows build tests)
turboderp committed
d88f185
T
Remove defunct test scripts
turboderp committed
cc601e5
T
qbench.py: Add more NVFP4 and FP8 formats
turboderp committed
ad0c971
T
CPU MoE: Fix Windows compile issues, add test build workflow
turboderp committed
abc78b9
T
qbench.py: Support NVFP4 (Transformers)
turboderp committed
6214054
July 24, 2026
T
qbench.py: Update histogram plot
turboderp committed
8882b0b
T
qbench.py: Add KLD histogram plots
turboderp committed
848897d
T
qbench.py: Fix some composite model quirks
turboderp committed
a9da393
T
longctx.py: Add extra-long test
turboderp committed
6b87433
T
SlidingAttention: Zero cache window before unstashing (kernel multiplies by zero to mask unused columns, but uninitialized values may be NaN)
turboderp committed
a0ef743
T
qbench.py: Improve plot
turboderp committed
b465c70
July 23, 2026
T
qbench.py: Properly hook output layer, accounting for multimodal reference models and logit softcapping
turboderp committed
4bf0165