[Bugfix] Fix wrong multi_modal_input format for CPU runner (#5451)

Isotr0py · web-flow · commit 2135cacb457b · 2024-06-12T16:20:18.000-07:00
diff --git a/vllm/worker/cpu_model_runner.py b/vllm/worker/cpu_model_runner.py
@@ -343,8 +343,8 @@ def execute_model(
             "kv_caches": kv_caches,
             "attn_metadata": attn_metadata,
         }
-        if self.vision_language_config:
-            execute_model_kwargs.update({"image_input": multi_modal_input})
+        if self.vision_language_config and multi_modal_input is not None:
+            execute_model_kwargs.update(multi_modal_input)
 
         hidden_states = model_executable(**execute_model_kwargs)
 

Original file line number	Diff line number	Diff line change
`@@ -343,8 +343,8 @@ def execute_model(`
`343`	`343`	`"kv_caches": kv_caches,`
`344`	`344`	`"attn_metadata": attn_metadata,`
`345`	`345`	`}`
`346`		`- if self.vision_language_config:`
`347`		`- execute_model_kwargs.update({"image_input": multi_modal_input})`
	`346`	`+ if self.vision_language_config and multi_modal_input is not None:`
	`347`	`+ execute_model_kwargs.update(multi_modal_input)`
`348`	`348`
`349`	`349`	`hidden_states = model_executable(**execute_model_kwargs)`
`350`	`350`