Set Weight Loading to be strict (#12151)

cccclai · web-flow · commit 1ccfdf4e00b7 · 2025-07-03T12:18:26.000-07:00
Summary: It often causes confusion and weird downstream error when the
weight is not properly loaded. Set `load_state_dict` to be True so the
error populated early.

Differential Revision: D77626023
diff --git a/examples/qualcomm/oss_scripts/llama/llama.py b/examples/qualcomm/oss_scripts/llama/llama.py
@@ -591,7 +591,7 @@ def permute(w, heads):
     for llama_instance in llama_instance_list:
         llama_instance.load_state_dict(
             state_dict,
-            strict=False,
+            strict=True,
             assign=True,
         )
     end_load_ts = time.time()

Original file line number	Diff line number	Diff line change
`@@ -591,7 +591,7 @@ def permute(w, heads):`
`591`	`591`	`for llama_instance in llama_instance_list:`
`592`	`592`	`llama_instance.load_state_dict(`
`593`	`593`	`state_dict,`
`594`		`- strict=False,`
	`594`	`+ strict=True,`
`595`	`595`	`assign=True,`
`596`	`596`	`)`
`597`	`597`	`end_load_ts = time.time()`