Skip to content

⚡️ Speed up function _quantize_k_cache_slow by 55% #341

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

⚡️ Speed up function _quantize_k_cache_slow by 55% #341

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 21 additions & 22 deletions python/sglang/srt/layers/attention/nsa/quant_k_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,37 +31,36 @@ def _quantize_k_cache_slow(
input_k_cache = input_k_cache.squeeze(2) # [num_blocks, block_size, d]
input_elem_size = input_k_cache.element_size()


# Precompute shapes
rope_len = input_elem_size * (d - dv)
result_shape = (num_blocks, block_size, dv + num_tiles * 4 + rope_len)
result = torch.empty(
(num_blocks, block_size, dv + num_tiles * 4 + input_elem_size * (d - dv)),
result_shape,
dtype=torch.float8_e4m3fn,
device=input_k_cache.device,
)
result_k_nope_part = result[..., :dv]
result_k_scale_factor = result[..., dv : dv + num_tiles * 4].view(torch.float32)
result_k_rope_part = result[..., dv + num_tiles * 4 :].view(input_k_cache.dtype)
result_k_rope_part[:] = input_k_cache[..., dv:]

for tile_idx in range(0, num_tiles):
cur_scale_factors_inv = (
torch.abs(
input_k_cache[..., tile_idx * tile_size : (tile_idx + 1) * tile_size]
)
.max(dim=-1)
.values
/ 448.0
) # [num_blocks, block_size]
# Use directly indexed assignment (avoid unnecessary slice copying).
result_k_rope_part[...] = input_k_cache[..., dv:]

# Vectorized quantization over tiles for significant speedup.
for tile_idx in range(num_tiles):
start = tile_idx * tile_size
end = (tile_idx + 1) * tile_size

input_tile = input_k_cache[..., start:end]
# Vectorize: torch.abs and .max
cur_scale_factors_inv = torch.abs(input_tile).amax(dim=-1) / 448.0 # Shape: [num_blocks, block_size]
result_k_scale_factor[:, :, tile_idx] = cur_scale_factors_inv

cur_scale_factors_inv.unsqueeze_(-1) # [num_blocks, block_size, 1]
cur_quantized_nope = (
input_k_cache[
..., tile_idx * tile_size : (tile_idx + 1) * tile_size
].float()
/ cur_scale_factors_inv.float()
).to(torch.float8_e4m3fn)
result_k_nope_part[..., tile_idx * tile_size : (tile_idx + 1) * tile_size] = (
cur_quantized_nope
)
# Instead of unsqueeze_: use broadcasting directly
cur_quantized_nope = (input_tile.float() / cur_scale_factors_inv.float().unsqueeze(-1)).to(torch.float8_e4m3fn)
result_k_nope_part[..., start:end] = cur_quantized_nope

# Fast (no copy) reshape to output format

result = result.view(num_blocks, block_size, 1, -1)
return result
Expand Down