add some cuTile kernels for Liger suites & Update mla autotune config & other updates

src/tilegym/ops/cutile/mla.py

@@ -34,6 +34,9 @@ def _mla_sm90_autotune_configs():
     for tm in [64, 128, 256]:
         for tn in [64, 128]:
             yield SimpleNamespace(TILE_M=tm, TILE_N=tn, num_ctas=1, occupancy=1)
+    for tm in [64, 128]:
+        for tn in [64, 128]:
+            yield SimpleNamespace(TILE_M=tm, TILE_N=tn, num_ctas=1, occupancy=2)
 
 
 @ct.kernel

src/tilegym/suites/liger/__init__.py

@@ -20,14 +20,32 @@
 # Import unified interface
 from .ops import cross_entropy
 from .ops import fused_linear_jsd
+from .ops import fused_neighborhood_attention
 from .ops import geglu
+from .ops import group_norm
 from .ops import jsd
+from .ops import kl_div
 from .ops import layer_norm
+from .ops import llama4_rope
+from .ops import multi_token_attention
+from .ops import qwen2vl_mrope
+from .ops import rope
+from .ops import sparsemax
+from .ops import tiled_mlp
 
 __all__ = [
     "cross_entropy",
     "fused_linear_jsd",
+    "fused_neighborhood_attention",
     "geglu",
+    "group_norm",
     "jsd",
+    "kl_div",
     "layer_norm",
+    "llama4_rope",
+    "multi_token_attention",
+    "qwen2vl_mrope",
+    "rope",
+    "sparsemax",
+    "tiled_mlp",
 ]

src/tilegym/suites/liger/cutile/__init__.py

@@ -6,24 +6,56 @@
 
 from . import cross_entropy  # noqa: F401
 from . import fused_linear_jsd  # noqa: F401
+from . import fused_neighborhood_attention  # noqa: F401
 from . import geglu  # noqa: F401
+from . import group_norm  # noqa: F401
 from . import jsd  # noqa: F401
+from . import kl_div  # noqa: F401
 from . import layer_norm  # noqa: F401
+from . import llama4_rope  # noqa: F401
+from . import multi_token_attention  # noqa: F401
+from . import qwen2vl_mrope  # noqa: F401
+from . import rope  # noqa: F401
+from . import sparsemax  # noqa: F401
+from . import tiled_mlp  # noqa: F401
 from .cross_entropy import CrossEntropyCuTileFunction  # noqa: F401
 from .fused_linear_jsd import FusedLinearJSDCuTileFunction  # noqa: F401
 from .geglu import GEGLUCuTileFunction  # noqa: F401
+from .group_norm import GroupNormCuTileFunction  # noqa: F401
 from .jsd import JSDCuTileFunction  # noqa: F401
+from .kl_div import KLDivCuTileFunction  # noqa: F401
 from .layer_norm import LayerNormCuTileFunction  # noqa: F401
+from .llama4_rope import Llama4RopeCuTileFunction  # noqa: F401
+from .multi_token_attention import MultiTokenAttentionCuTileFunction  # noqa: F401
+from .qwen2vl_mrope import Qwen2VLMRopeCuTileFunction  # noqa: F401
+from .rope import RopeCuTileFunction  # noqa: F401
+from .sparsemax import SparsemaxCuTileFunction  # noqa: F401
 
 __all__ = [
     "CrossEntropyCuTileFunction",
     "FusedLinearJSDCuTileFunction",
     "GEGLUCuTileFunction",
+    "GroupNormCuTileFunction",
     "JSDCuTileFunction",
+    "KLDivCuTileFunction",
     "LayerNormCuTileFunction",
+    "Llama4RopeCuTileFunction",
+    "MultiTokenAttentionCuTileFunction",
+    "SparsemaxCuTileFunction",
     "cross_entropy",
     "fused_linear_jsd",
+    "fused_neighborhood_attention",
     "geglu",
+    "group_norm",
     "jsd",
+    "kl_div",
     "layer_norm",
+    "llama4_rope",
+    "multi_token_attention",
+    "Qwen2VLMRopeCuTileFunction",
+    "RopeCuTileFunction",
+    "qwen2vl_mrope",
+    "rope",
+    "sparsemax",
+    "tiled_mlp",
 ]