Change optimizations for Gemma3ForCausalLM vLLM to accuracy

MohammedTaherMcW · MohammedTaherMcW · commit e324df80cd8c · 2025-08-05T15:20:11.000Z
diff --git a/models/tt_transformers/tt/generator_vllm.py b/models/tt_transformers/tt/generator_vllm.py
@@ -391,7 +391,7 @@ def initialize_vllm_model(
             max_seq_len=max_seq_len,
             n_layers=n_layers,
             dtype=ttnn.bfloat16,
-            optimizations=DecodersPrecision.performance,
+            optimizations=DecodersPrecision.accuracy,
         )
         return cls(tt_model, model_args, mesh_device)
 

Original file line number	Diff line number	Diff line change
`@@ -391,7 +391,7 @@ def initialize_vllm_model(`
`391`	`391`	`max_seq_len=max_seq_len,`
`392`	`392`	`n_layers=n_layers,`
`393`	`393`	`dtype=ttnn.bfloat16,`
`394`		`- optimizations=DecodersPrecision.performance,`
	`394`	`+ optimizations=DecodersPrecision.accuracy,`
`395`	`395`	`)`
`396`	`396`	`return cls(tt_model, model_args, mesh_device)`
`397`	`397`