biondizzle
f3b551956d
- Move dead dsv4/ modules to dsv4/_archive/ (52 files)
- model/{dsv4,mtp,layer,layer_schedule}
- layers/{embedding,attention,ffn,norm} (kept linear,mhc,router,moe,shared_expert,grouped_linear - live)
- cache/*, kernels/cache/*, kernels/indexer/{csa_indexer,score_topk,compute_valid_lens}
- kernels/router/{nvfp4_fused_router,dense_router_decode_kernel,dense_router_prefill}
- ops/{topk,topk_select,rope,router}, loader/{hf_checkpoint,layout_convert}
- reference/{attention,compressor,csa_attention,moe_pipeline}
- kernels/compressor/{compress_tail,csa_hca}
- Restore dsv4/ops/{router,custom_ops}.py (needed by live layers)
- Fix dsv4/kernels/{indexer,compressor,attention}/__init__.py (removed broken imports)
- Remove preload_all() from loader.py (dead, referenced nonexistent .cu file)
- Fix loader.py docstring (fused_amax_quantize_nvfp4 → quantize_nvfp4_from_buffer)
- Move broken tests to tests/e2e_archive/
- test_fused_router, production_values_test, e2e/{one_layer,model_construction,csa_hca}
- vLLM has 0 imports of dsv4 (Step 0 confirmed)
45 lines
1.5 KiB
Python
45 lines
1.5 KiB
Python
"""Python wrapper for the topk_select CUDA kernel.
|
|
|
|
Lazy-loads the topk_select extension (same pattern as dsv4/ops/topk.py).
|
|
This is the general top-k primitive reused by the router and the CSA indexer.
|
|
"""
|
|
|
|
import os
|
|
import torch
|
|
|
|
_kernel_module = None
|
|
|
|
|
|
def _get_kernel_module():
|
|
"""Lazy-load the topk_select CUDA extension."""
|
|
global _kernel_module
|
|
if _kernel_module is not None:
|
|
return _kernel_module
|
|
|
|
from torch.utils.cpp_extension import load
|
|
kernel_dir = os.path.join(os.path.dirname(__file__), "kernels", "cuda")
|
|
_kernel_module = load(
|
|
name="topk_select",
|
|
sources=[os.path.join(kernel_dir, "topk_select.cu")],
|
|
extra_cuda_cflags=["-O3", "--generate-code=arch=arch=compute_100a,code=[sm_100a]"],
|
|
verbose=False,
|
|
)
|
|
return _kernel_module
|
|
|
|
|
|
def topk_select(
|
|
scores: torch.Tensor, # [num_rows, E] float32, row-major contiguous
|
|
k: int, # number to select (currently only k=6 supported)
|
|
) -> tuple[torch.Tensor, torch.Tensor]:
|
|
"""Select top-k indices and values from each row of scores.
|
|
|
|
Returns (values, indices) where:
|
|
values: [num_rows, k] float32 — top-k scores in descending order
|
|
indices: [num_rows, k] int32 — top-k indices (0-based, lower index wins on ties)
|
|
|
|
One block per row, 64 threads per block, per-thread register min-heap
|
|
with shared-memory merge. O(E * log k) per row.
|
|
"""
|
|
mod = _get_kernel_module()
|
|
return mod.topk_select(scores, k)
|