Repository navigation

All pull requests

Labels (17)Milestones (0)New pull request

Open

(46)

Closed

(3,020)

Fix premature rounding in addmm
#4769tetsuo opened Oct 25, 2026
Check scatter and gather indices are in bounds
#4768turingH opened Oct 25, 2026
[jaccl] Detect dead RDMA links instead of crashing
#4767ipazarbasi opened Oct 25, 2026
Route of per-channel gate in gated_delta_update fixed
#4766Exorust opened Oct 25, 2026
Fix gguf tensor to f16 size overflow
#4764turingH opened Oct 25, 2026
Fix GPU concatenate for outputs larger than 2^31 elements
#4762citruslab opened Oct 24, 20263 tasks done
[CUDA] Expose compilation targets via device_info
#4761dhiltgen opened Oct 24, 2026Contributor
Keep g and beta in float32 in gated_delta_update
#4760jvmenen opened Oct 24, 2026
Fix GPU concatenate for outputs larger than 2^31 elements
#4758tivris opened Oct 24, 2026Contributor
Add optional source IP binding for ring connections
#4757ronaldmannak opened Oct 24, 2026Contributor
Fix FP8 NaN preservation in to_fp8 and E8M0 scale dequantization
#4756FaizPalwala opened Oct 23, 2026
Fix any/all for int8 and uint8 on Metal
#4755MarcosAsh opened Oct 23, 2026Contributor
Bigger MoE tiles on M1 GPUs
#4754a3rosol opened Oct 23, 2026
1
Fix flipped bounds check in the Metal large-filter convolution loader
#4750ssmall256 opened Oct 23, 2026Contributor
Fix vmap of tri_inv ignoring the triangular flags
#4748mahimn01 opened Oct 22, 2026
Fixes MLX/ErrorHandler.swift:345: Fatal error due to missing stream argument
#4746Satchitananda opened Oct 22, 20261 task done
1
[CUDA] Avoid multi-root graphs on DGX/RTX spark
#4745dhiltgen opened Oct 22, 2026Contributor
3
Use gemv_wide for fp32 matmuls with 2-3 rows
#4742utsav1033 opened Oct 22, 2026
Cleanup metal quantized kernel instantiations
#4741zcbenz opened Oct 22, 2026Member
feat(nn): add freqs and partial_rotary_factor support to RoPE
#4732shadyuwugurl opened Oct 20, 2026
[Metal] Add SDPA fast paths for speculative decoding
#4723dhiltgen opened Oct 18, 2026Contributor
3
Add optional int8 path for quantized matmul with affine weights
#4722magnuslundstedt opened Oct 18, 2026Contributor
1
Fix silent Metal grid truncation in softmax, logsumexp, rms_norm and layer_norm past 2^27 rows
#4717ArupKantiDas opened Oct 17, 2026
4
Fix sharded-to-all bias and integer split positions on quantized layers
#4715ronaldmannak opened Oct 17, 2026Contributor
Debug time graph verification
#4711zcbenz opened Oct 16, 2026Member

Navigation Menu

Repository navigation

All pull requests

Fix premature rounding in addmm

Check scatter and gather indices are in bounds

[jaccl] Detect dead RDMA links instead of crashing

Route of per-channel gate in gated_delta_update fixed

Fix gguf tensor to f16 size overflow

Fix GPU concatenate for outputs larger than 2^31 elements

[CUDA] Expose compilation targets via device_info

Keep `g` and `beta` in float32 in `gated_delta_update`

Fix GPU concatenate for outputs larger than 2^31 elements

Add optional source IP binding for ring connections

Fix FP8 NaN preservation in to_fp8 and E8M0 scale dequantization

Fix any/all for int8 and uint8 on Metal

Bigger MoE tiles on M1 GPUs

Fix flipped bounds check in the Metal large-filter convolution loader

Fix vmap of tri_inv ignoring the triangular flags

Fixes MLX/ErrorHandler.swift:345: Fatal error due to missing stream argument

[CUDA] Avoid multi-root graphs on DGX/RTX spark

Use gemv_wide for fp32 matmuls with 2-3 rows

Cleanup metal quantized kernel instantiations

feat(nn): add freqs and partial_rotary_factor support to RoPE

[Metal] Add SDPA fast paths for speculative decoding

Add optional int8 path for quantized matmul with affine weights

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

Debug time graph verification

Pull requests