fix awq kernel (#3618)

grimoire · web-flow · commit 9617d7d841f8 · 2025-06-09T20:43:22.000+08:00
* fix awq kernel

* reset to zero
diff --git a/lmdeploy/pytorch/kernels/cuda/awq_kernels.py b/lmdeploy/pytorch/kernels/cuda/awq_kernels.py
@@ -85,6 +85,7 @@ def _unpack_weight(weight):
 @triton.autotune(
     configs=get_cuda_autotune_config(),
     key=['N', 'K'],
+    reset_to_zero=['c_ptr'],
 )
 @triton.jit
 def awq_linear_kernel(

Original file line number	Diff line number	Diff line change
`@@ -85,6 +85,7 @@ def _unpack_weight(weight):`
`85`	`85`	`@triton.autotune(`
`86`	`86`	`configs=get_cuda_autotune_config(),`
`87`	`87`	`key=['N', 'K'],`
	`88`	`+ reset_to_zero=['c_ptr'],`
`88`	`89`	`)`
`89`	`90`	`@triton.jit`
`90`	`91`	`def awq_linear_kernel(`