bug fix

chichun-charlie-liu · chichun-charlie-liu · commit 68de6c1eec2e · 2025-06-27T04:00:51.000Z
Signed-off-by: cliu-us &lt;cliu@us.ibm.com&gt;
diff --git a/fms_mo/custom_ext_kernels/utils.py b/fms_mo/custom_ext_kernels/utils.py
@@ -938,7 +938,7 @@ def lower_qmodel_triton(
         else:
             new_lin = LinearFPxAcc.from_nn(
                 m,
-                trun_bits=max_acc_bits,
+                trun_bits=num_lsb_to_truncate,
                 chunk_size=chunk_size,
                 dynamic_fp8=linFP_dyn_code,
                 clamp_acc_to_dl16=clamp_acc_to_dl16,
diff --git a/fms_mo/modules/linear.py b/fms_mo/modules/linear.py
@@ -2047,7 +2047,7 @@ def from_nn(cls, nnlin, trun_bits=0, **kwargs):
             nnlin.in_features,
             nnlin.out_features,
             bias=nnlin.bias is not None,
-            device=target_device,
+            device="meta",
         )
 
         lin24acc.weight = nnlin.weight

Original file line number	Diff line number	Diff line change
`@@ -2047,7 +2047,7 @@ def from_nn(cls, nnlin, trun_bits=0, **kwargs):`
`2047`	`2047`	`nnlin.in_features,`
`2048`	`2048`	`nnlin.out_features,`
`2049`	`2049`	`bias=nnlin.bias is not None,`
`2050`		`- device=target_device,`
	`2050`	`+ device="meta",`
`2051`	`2051`	`)`
`2052`	`2052`
`2053`	`2053`	`lin24acc.weight = nnlin.weight`