[Quantization] Support native Quark W4A16 INT4/UINT4 exports in vLLM (#48606)

Signed-off-by: HongWei Meng <HongWei.Meng@amd.com>
This commit is contained in:
meng-amd
2026-09-18 17:52:46 +08:00
committed by GitHub
parent 67a8a3f926
commit 6df2b1a8c6
9 changed files with 1124 additions and 6 deletions
@@ -93,7 +93,7 @@ def _replace_or_register_parameter(
def _convert_awq_to_standard_format(
layer: torch.nn.Module,
w_q_name: str,
w_zp_name: str,
w_zp_name: str | None,
size_bits: int,
) -> None:
"""Convert AWQ weight and zero-point tensors to standard GPTQ-like format.
@@ -101,6 +101,7 @@ def _convert_awq_to_standard_format(
AWQ packs qweight along the output dim with a non-standard bit order.
This converts to standard bit order and repacks qweight along the input
dim, matching the format expected by the MPLinearKernel framework.
If w_zp_name is None (symmetric quantization), only the weight is converted.
"""
pack_factor = 32 // size_bits
mask = (1 << size_bits) - 1
@@ -139,6 +140,9 @@ def _convert_awq_to_standard_format(
)
setattr(layer, w_q_name, new_param)
if w_zp_name is None:
return
# --- Convert qzeros: fix AWQ bit ordering and repack
# AWQ qzeros: (G, N // pack) packed along dim 1, AWQ bit order
# Target: (N // pack, G) packed along dim 0, standard bit order