[Refactor] kernels: Register agent-generated kernels under their KDA / Cake backend (#45277)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
fdd89d182a
commit
0788f1fdd0
@@ -104,16 +104,17 @@ and will keep growing; they are implementation homes, not operator groups.
|
||||
(Humanize2) workflow, in CUDA, Triton, or CuTe DSL, with their JIT sources in
|
||||
`kda_kernels/csrc/`. `KernelBackend.KDA` records that provenance.
|
||||
- `cake_kernels/` — thin adapters to Cake-generated kernels that FlashInfer
|
||||
distributes (generated sources and caches stay in FlashInfer); they register
|
||||
with the provenance of that library (`KernelBackend.FLASHINFER`).
|
||||
distributes (generated sources and caches stay in FlashInfer).
|
||||
`KernelBackend.CAKE` records that provenance.
|
||||
|
||||
Their public ops are registered into the logical groups with
|
||||
`KernelSpec(op="<group>.<name>", backend=..., target="sglang.kernels.kda_kernels.<mod>:<fn>")`
|
||||
(or a `cake_kernels` target), and runtime code imports the `ops.<group>`
|
||||
facade, never these packages. Put a new agent-generated kernel in the matching
|
||||
directory with its provenance (see `kda_kernels/README.md`). List them from the
|
||||
registry by target (an agent kernel that is the op's only implementation may
|
||||
keep its language backend, e.g. `TRITON`):
|
||||
directory and register it with that directory's backend (`KDA` or `CAKE`),
|
||||
whatever language it is written in (see `kda_kernels/README.md`);
|
||||
`test_kernels_namespace.py` enforces the pairing. List them from the registry
|
||||
by backend:
|
||||
|
||||
```python
|
||||
import sglang.kernels.ops # populate the registry
|
||||
@@ -122,10 +123,7 @@ from sglang.kernels import KernelBackend, registry
|
||||
agent_kernels = [
|
||||
spec
|
||||
for spec in registry.all_specs()
|
||||
if spec.backend is KernelBackend.KDA
|
||||
or spec.target.startswith(
|
||||
("sglang.kernels.kda_kernels.", "sglang.kernels.cake_kernels.")
|
||||
)
|
||||
if spec.backend in (KernelBackend.KDA, KernelBackend.CAKE)
|
||||
]
|
||||
```
|
||||
|
||||
|
||||
@@ -109,6 +109,7 @@ BACKEND_METHODS: Dict[KernelBackend, str] = {
|
||||
KernelBackend.FLYDSL: "forward_flydsl",
|
||||
KernelBackend.HELION: "forward_helion",
|
||||
KernelBackend.KDA: "forward_kda",
|
||||
KernelBackend.CAKE: "forward_cake",
|
||||
KernelBackend.FLASHINFER: "forward_flashinfer",
|
||||
KernelBackend.DEEPGEMM: "forward_deepgemm",
|
||||
KernelBackend.AITER: "forward_aiter",
|
||||
@@ -124,6 +125,7 @@ _METHOD_BACKEND_LABELS: Dict[str, str] = {
|
||||
# explicitly when wanted. Per-op priority overrides this (see BaseFusedOp).
|
||||
DEFAULT_PRIORITY: Tuple[KernelBackend, ...] = (
|
||||
KernelBackend.KDA,
|
||||
KernelBackend.CAKE,
|
||||
KernelBackend.AOT,
|
||||
KernelBackend.JIT,
|
||||
KernelBackend.FLASHINFER,
|
||||
@@ -428,6 +430,9 @@ class BaseFusedOp(nn.Module, ABC):
|
||||
def forward_kda(self, *args, **kwargs):
|
||||
raise NotImplementedError(f"{self._op_label()}: no KDA backend")
|
||||
|
||||
def forward_cake(self, *args, **kwargs):
|
||||
raise NotImplementedError(f"{self._op_label()}: no Cake backend")
|
||||
|
||||
def forward_flashinfer(self, *args, **kwargs):
|
||||
raise NotImplementedError(f"{self._op_label()}: no flashinfer backend")
|
||||
|
||||
|
||||
@@ -75,7 +75,7 @@ del _mod, _fn
|
||||
register_kernel(
|
||||
KernelSpec(
|
||||
op="attention.kda_qwen38_qsa_sm121",
|
||||
backend=KernelBackend.TRITON,
|
||||
backend=KernelBackend.KDA,
|
||||
target=("sglang.kernels.kda_kernels.qwen38_qsa_sm121:qwen38_qsa_sm121"),
|
||||
capabilities=frozenset(
|
||||
{CapabilityRequirement.cuda(min_sm=(12, 1), max_sm=(12, 1))}
|
||||
@@ -123,7 +123,7 @@ def qwen38_qsa_sm121_varlen(
|
||||
"unsupported SM121 QSA call: expected BF16 D=256, 12:1 GQA, "
|
||||
"TP1 24Q/2KV or TP2 12Q/1KV, bs<=128, and selected KV<=2055"
|
||||
)
|
||||
return get_kernel("attention.kda_qwen38_qsa_sm121", KernelBackend.TRITON)(
|
||||
return get_kernel("attention.kda_qwen38_qsa_sm121", KernelBackend.KDA)(
|
||||
q,
|
||||
k,
|
||||
v,
|
||||
|
||||
@@ -149,7 +149,7 @@ _SPECS: tuple[tuple[str, KernelBackend, str, frozenset], ...] = (
|
||||
),
|
||||
(
|
||||
"diffusion.scale_residual_norm_scale_shift_nvfp4",
|
||||
KernelBackend.JIT,
|
||||
KernelBackend.KDA,
|
||||
"sglang.kernels.kda_kernels.norm_scale_shift_jit:try_fused_scale_residual_norm_scale_shift_nvfp4",
|
||||
_CUDA,
|
||||
),
|
||||
@@ -173,13 +173,13 @@ _SPECS: tuple[tuple[str, KernelBackend, str, frozenset], ...] = (
|
||||
),
|
||||
(
|
||||
"diffusion.layernorm_modulate",
|
||||
KernelBackend.TRITON,
|
||||
KernelBackend.KDA,
|
||||
"sglang.kernels.kda_kernels.layernorm_modulate_triton:fused_layernorm_modulate",
|
||||
_CUDA,
|
||||
),
|
||||
(
|
||||
"diffusion.qk_head_layernorm",
|
||||
KernelBackend.TRITON,
|
||||
KernelBackend.KDA,
|
||||
"sglang.kernels.kda_kernels.layernorm_modulate_triton:fused_qk_head_layernorm",
|
||||
_CUDA,
|
||||
),
|
||||
|
||||
@@ -19,7 +19,7 @@ if TYPE_CHECKING:
|
||||
register_kernel(
|
||||
KernelSpec(
|
||||
op="sampling.softmax",
|
||||
backend=KernelBackend.FLASHINFER,
|
||||
backend=KernelBackend.CAKE,
|
||||
target="sglang.kernels.cake_kernels.sampling:softmax",
|
||||
capabilities=frozenset(
|
||||
{
|
||||
@@ -32,7 +32,7 @@ register_kernel(
|
||||
|
||||
def cake_softmax(logits: torch.Tensor) -> torch.Tensor:
|
||||
"""Explicit Cake softmax entry point for Blackwell qualification."""
|
||||
return get_kernel("sampling.softmax", KernelBackend.FLASHINFER)(logits)
|
||||
return get_kernel("sampling.softmax", KernelBackend.CAKE)(logits)
|
||||
|
||||
|
||||
def softmax(
|
||||
|
||||
@@ -46,6 +46,7 @@ class KernelBackend(str, Enum):
|
||||
FLYDSL = "flydsl" # FlyDSL MLIR compiler (device=HIP, gfx950)
|
||||
HELION = "helion"
|
||||
KDA = "KDA" # Kernel Design Agents generated implementation
|
||||
CAKE = "cake" # Cake-generated implementation, distributed by FlashInfer
|
||||
FLASHINFER = "flashinfer"
|
||||
DEEPGEMM = "deepgemm"
|
||||
AITER = "aiter" # AMD aiter library (device=HIP)
|
||||
|
||||
Reference in New Issue
Block a user