[Attention Backend] TurboQuant: 2-bit KV cache compression with 4x capacity (#38479)

Signed-off-by: vibhavagarwal5 <vibhavagarwal5@gmail.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Xinyu Chen <xinyu1.chen@intel.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Vibhav Agarwal
2026-04-14 19:57:13 -07:00
committed by GitHub
co-authored by Xinyu Chen Michael Goin
parent 3bfe55a037
commit f4b42df048
27 changed files with 2963 additions and 3 deletions
@@ -0,0 +1,24 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""TurboQuant quantizer utilities.
Serving path uses generate_wht_signs() for WHT rotation sign buffers.
Triton kernels handle all quantization, packing, and dequantization on GPU.
"""
import torch
_CPU = torch.device("cpu")
def generate_wht_signs(d: int, seed: int, device: torch.device = _CPU) -> torch.Tensor:
"""Generate deterministic random ±1 signs for WHT rotation.
Used with Walsh-Hadamard Transform for per-layer rotation randomization.
Same seed derivation as QR (per-layer via seed + layer_idx * stride).
"""
gen = torch.Generator(device="cpu")
gen.manual_seed(seed)
bits = torch.randint(0, 2, (d,), generator=gen, device="cpu")
signs = bits.float() * 2 - 1
return signs.to(device)