Repository navigation

All pull requests

Fix silent Metal grid truncation in softmax, logsumexp, rms_norm and layer_norm past 2^27 rows
#4717ArupKantiDas opened Oct 17, 2026
Fix sharded-to-all bias and integer split positions on quantized layers
#4715ronaldmannak opened Oct 17, 2026Contributor
Allow configuring the CPU JIT cache directory
#4714cerisier opened Oct 16, 2026
Make broadcast_arrays return a tuple
#4713prady0t opened Oct 16, 2026Contributor
[Metal] Tile fp quantized_matmul for 6-16 rows
#4712dhiltgen opened Oct 16, 2026Contributor
Debug time graph verification
#4711zcbenz opened Oct 16, 2026Member
Release the previous value through ~array() on assignment
#4710yangzichao opened Oct 16, 20262 of 4 tasks
[Metal] Use half tile for fp4 qmm
#4709dhiltgen opened Oct 16, 2026Contributor
[Metal] Fix SDPA f16/f32 d512 max threadgroup for M1/M2 generation
#4708dhiltgen opened Oct 15, 2026Contributor
Fix Windows CUDA 13 DLL loading
#4706luciusacarr opened Oct 15, 2026
Fix fftn size one axis aliasing a temporary
#4705StockerMC opened Oct 15, 2026
1
Make GPU scheduler know the spin is waiting for something
#4703aleroot opened Oct 15, 2026Contributor
3
Fix fast.cross_entropy VJP on non-contiguous Metal inputs
#4702StockerMC opened Oct 15, 2026
Fixed lshift and rshift to return 0 for unsigned and -1 for signed value when shift is greater than dtype size
#4694aaishwarymishra opened Oct 14, 2026Contributor
Simplify docs generation by ditching doxygen
#4690zcbenz opened Oct 14, 2026Member
Fix negative padding with input_dilation > 1 in conv_general
#4688MarcosAsh opened Oct 13, 2026Contributor
2
Speed up transposed quantized matmuls on M5
#4682wyanzhao opened Oct 11, 2026Contributor
Add "config" module
#4647zcbenz opened Oct 8, 2026Member
2
Fix custom_function vmap and jvp traversal with opaque bodies
#4612axiom-of-choice opened Oct 2, 2026Contributor1 task done
2
Perf: Metal kernel source hash cache
#4568STRML opened Sep 26, 2026
3
Support uneven tensor parallel sharding in distributed linear layers
#4465twallgren opened Sep 4, 2026
Fix strided_scan reading/writing out-of-bounds
#4430zcbenz opened Aug 30, 2026Member
3
Small input linear and cubic upsampling should use dense matmuls
#4418slink opened Aug 28, 2026
1
[CUDA] Grouped mm
#4390nastya236 opened Aug 24, 2026Collaborator
1
Add nanobind 3.0 split-mode support
#4359XXXXRT666 opened Aug 19, 2026Contributor4 tasks done
2

Navigation Menu

Repository navigation

All pull requests

Fix silent Metal grid truncation in softmax, logsumexp, rms_norm and layer_norm past 2^27 rows

Fix sharded-to-all bias and integer split positions on quantized layers

Allow configuring the CPU JIT cache directory

Make `broadcast_arrays` return a tuple

[Metal] Tile fp quantized_matmul for 6-16 rows

Debug time graph verification

Release the previous value through ~array() on assignment

[Metal] Use half tile for fp4 qmm

[Metal] Fix SDPA f16/f32 d512 max threadgroup for M1/M2 generation

Fix Windows CUDA 13 DLL loading

Fix fftn size one axis aliasing a temporary

Make GPU scheduler know the spin is waiting for something

Fix fast.cross_entropy VJP on non-contiguous Metal inputs

Fixed lshift and rshift to return 0 for unsigned and -1 for signed value when shift is greater than dtype size

Simplify docs generation by ditching doxygen

Fix negative padding with input_dilation > 1 in conv_general

Speed up transposed quantized matmuls on M5

Add "config" module

Fix custom_function vmap and jvp traversal with opaque bodies

Perf: Metal kernel source hash cache

Support uneven tensor parallel sharding in distributed linear layers

Fix strided_scan reading/writing out-of-bounds

Small input linear and cubic upsampling should use dense matmuls

[CUDA] Grouped mm

Add nanobind 3.0 split-mode support

Pull requests