77 lines
1.9 KiB
Python
77 lines
1.9 KiB
Python
"""Embedding kernels."""
|
|
|
|
from sglang.kernels.registry import register_kernel
|
|
from sglang.kernels.spec import CUDA_ONLY, KernelBackend, KernelSpec
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.vocab_parallel_embedding",
|
|
backend=KernelBackend.TRITON,
|
|
target=(
|
|
"sglang.kernels.ops.embeddings.vocab_parallel_embedding:"
|
|
"vocab_parallel_embedding"
|
|
),
|
|
)
|
|
)
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.engram_gather",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.engram_gather:engram_gather",
|
|
)
|
|
)
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.engram_gather_peer",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.engram_gather:engram_gather_peer",
|
|
)
|
|
)
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.engram_hash_ids",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.engram_hash:engram_hash_ids",
|
|
)
|
|
)
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.engram_hash_ids_and_commit",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.engram_hash:engram_hash_ids_and_commit",
|
|
)
|
|
)
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.engram_commit_history",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.engram_hash:engram_commit_history",
|
|
)
|
|
)
|
|
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.fused_engram_gate",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.engram_gate:fused_engram_gate",
|
|
)
|
|
)
|
|
|
|
__all__ = []
|
|
|
|
|
|
# Public entry points inventoried by logical operator group (RFC #29630).
|
|
register_kernel(
|
|
KernelSpec(
|
|
op="embeddings.fused_qwen4_ngram_hash",
|
|
backend=KernelBackend.TRITON,
|
|
target="sglang.kernels.ops.embeddings.qwen4_ngram:fused_qwen4_ngram_hash",
|
|
capabilities=CUDA_ONLY,
|
|
)
|
|
)
|