[NPU] recover accuracy for gemma3-4b-it from 54% to 72% (reduced by transformer5.3) (#21537)

2026-05-13 16:46:04 +08:00
parent d8f7b78a29
commit 9d0be860a4
1 changed files with 2 additions and 1 deletions
@@ -578,8 +578,9 @@ class Gemma3TextModel(PreTrainedModel):
        global_config = copy.deepcopy(config)
        global_config.rope_parameters = {
            "rope_type": "default",
            "rope_theta": global_theta,
            "factor": config.rope_parameters["full_attention"]["factor"],
            "rope_type": "linear",
        }
        self.rotary_emb = Gemma3RotaryEmbedding(config=global_config)
        self.gradient_checkpointing = False