[ROCm][FEAT] Integrate aiter gemm w8a8 ptpc (#33773)

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
This commit is contained in:
vllmellm
2026-04-16 09:55:29 +08:00
committed by GitHub
parent 343f65234b
commit 5f7fab881a
9 changed files with 279 additions and 23 deletions
@@ -397,8 +397,6 @@ class Fp8LinearMethod(LinearMethodBase):
if self.block_quant:
assert not self.act_q_static
self.fp8_linear.process_weights_after_loading(layer)
# If checkpoint not serialized fp8, quantize the weights.
else:
# If checkpoint is fp8 per-tensor, handle that there are N scales for N
@@ -428,6 +426,8 @@ class Fp8LinearMethod(LinearMethodBase):
else:
layer.input_scale = None
self.fp8_linear.process_weights_after_loading(layer)
def apply(
self,
layer: torch.nn.Module,