mirror of
https://github.com/FlashML-org/FreeToken.git
synced 2026-11-22 14:57:29 +08:00
* refactor(quant): config, scheme and method layers for checkpoint quantization QuantConfig parses the checkpoint's quantization_config into a QuantScheme per projection; every Linear*, ParallelLMHead and MoE layer takes quant_config and a prefix, declares its tensors through a QuantMethod and selects its kernel at construction. MoE kernels own the expert bank layout, pack and apply, so the offload cache, the FTW converter and the CPU executor follow one layout. - --quant-backend layer[.kind]=name selects kernels (--nvfp4-backend forwards to it) - --moe-strategy replaces --moe-backend; the old spelling stays as an alias - --moe-cpu-layers auto for Windows/WSL pin budgets; --moe-cache-auto respects the kernel's slot limit - removes the per-family quantized layer classes, bank providers, format env switches and GLM runtime fp8
30 lines
1.1 KiB
Python
30 lines
1.1 KiB
Python
"""Static window / compressed top-k index builders."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
|
|
def get_window_topk_idxs(window_size, bsz, seqlen, start_pos):
|
|
if start_pos >= window_size - 1:
|
|
start_pos %= window_size
|
|
matrix = torch.cat([torch.arange(start_pos + 1, window_size), torch.arange(0, start_pos + 1)], dim=0)
|
|
elif start_pos > 0:
|
|
matrix = F.pad(torch.arange(start_pos + 1), (0, window_size - start_pos - 1), value=-1)
|
|
else:
|
|
base = torch.arange(seqlen).unsqueeze(1)
|
|
matrix = (base - window_size + 1).clamp(0) + torch.arange(min(seqlen, window_size))
|
|
matrix = torch.where(matrix > base, -1, matrix)
|
|
return matrix.unsqueeze(0).expand(bsz, -1, -1)
|
|
|
|
|
|
def get_compress_topk_idxs(ratio, bsz, seqlen, start_pos, offset):
|
|
if start_pos > 0:
|
|
matrix = torch.arange(0, (start_pos + 1) // ratio) + offset
|
|
else:
|
|
matrix = torch.arange(seqlen // ratio).repeat(seqlen, 1)
|
|
mask = matrix >= torch.arange(1, seqlen + 1).unsqueeze(1) // ratio
|
|
matrix = torch.where(mask, -1, matrix + offset)
|
|
return matrix.unsqueeze(0).expand(bsz, -1, -1)
|