Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(22)

Closed

(6)

[BugFix][Multi-GPU] Guard the last two unguarded MoE launch sites
#40PerryLink opened Oct 14, 2026
[BugFix][Multi-GPU] Make MoE device resolution follow the input tensor
#39PerryLink opened Oct 14, 2026
fix(engram): fit sinkhorn step_reduce into smaller shared-memory budgets (SM120)
#37xzwgit opened Oct 1, 2026
fix(config): validate token alignment before updating global state
#35Afloat16 opened Oct 1, 2026
Extend CUDA Rotary Embedding Dtypes and Dimensions
#33CharlesXu-HQ opened Sep 25, 2026
Support Group-128 Fused Add RMSNorm FP8 Cast
#32CharlesXu-HQ opened Sep 25, 2026
Add Fused RMSNorm FP8 Quant Kernel
#31CharlesXu-HQ opened Sep 25, 2026
[BugFix][Multi-GPU] Allocate and launch MoE outputs on the input device
#30PerryLink opened Sep 17, 2026
6
fix(testing): support fp16 and float8_e5m2 in dtype_to_str
#29LeonxLJX opened Sep 1, 2026
6
[BugFix][quant] Fix per_channel_cast_fused expand tails mishandle bug
#27GenTang opened Jul 28, 2026
4
[BugFix][Compatibility] Fix tilelang import bug
#25GenTang opened Jul 26, 2026
6
[BugFix][mhc] Apply RMSNorm weight fusion in the no-grad mhc_pre path
#23GenTang opened Jul 26, 2026
15
[BugFix][MoE] Guard fused mapping edge tiles
#20JayceSu98 opened Jun 12, 2026
[BugFix][Engram] Support SM counts not divisible by four
#17JayceSu98 opened May 28, 2026
[BugFix][Quant] Guard scale tails and FP32 absmax
#15JayceSu98 opened May 28, 2026
[AMD] Add MI350 (gfx950 / HIP) support for TileKernels operators
#13zhangnju opened Apr 25, 2026
fix(tests/pytest_random_plugin.py): also seed Python's random
#12yurekami opened Apr 25, 2026
fix(mhc): MHCPreNormFn.backward returns wrong number of gradients
#10yurekami opened Apr 25, 2026
fix(testing/bench.py): make_param_key use _format_value for non-width keys
#9pandacooming opened Apr 24, 2026
fix(testing/bench.py): add dict support to _format_value
#7pandacooming opened Apr 24, 2026
fix(testing/bench.py): add float16 and float8_e5m2 to dtype_to_str mapping
#5pandacooming opened Apr 24, 2026
Add Code Format Check CI (yapf + ruff)
#3pandacooming opened Apr 24, 2026

Navigation Menu

Repository navigation

All pull requests

[BugFix][Multi-GPU] Guard the last two unguarded MoE launch sites

[BugFix][Multi-GPU] Make MoE device resolution follow the input tensor

fix(engram): fit sinkhorn step_reduce into smaller shared-memory budgets (SM120)

fix(config): validate token alignment before updating global state

Extend CUDA Rotary Embedding Dtypes and Dimensions

Support Group-128 Fused Add RMSNorm FP8 Cast

Add Fused RMSNorm FP8 Quant Kernel

[BugFix][Multi-GPU] Allocate and launch MoE outputs on the input device

fix(testing): support fp16 and float8_e5m2 in dtype_to_str

[BugFix][quant] Fix per_channel_cast_fused expand tails mishandle bug

[BugFix][Compatibility] Fix tilelang import bug

[BugFix][mhc] Apply RMSNorm weight fusion in the no-grad mhc_pre path

[BugFix][MoE] Guard fused mapping edge tiles

[BugFix][Engram] Support SM counts not divisible by four

[BugFix][Quant] Guard scale tails and FP32 absmax

[AMD] Add MI350 (gfx950 / HIP) support for TileKernels operators

fix(tests/pytest_random_plugin.py): also seed Python's `random`

fix(mhc): MHCPreNormFn.backward returns wrong number of gradients

fix(testing/bench.py): make_param_key use _format_value for non-width keys

fix(testing/bench.py): add dict support to _format_value

fix(testing/bench.py): add float16 and float8_e5m2 to dtype_to_str mapping

Add Code Format Check CI (yapf + ruff)

Pull requests