From 11fd7ea639cf3c4fae29322d8e5c839ff6f8a1ca Mon Sep 17 00:00:00 2001 From: Patrick von Platen Date: Tue, 19 Nov 2024 18:33:06 +0100 Subject: [PATCH] [Pixtral-Large] Pixtral actually has no bias in vision-lang adapter (#10449) --- vllm/model_executor/models/pixtral.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/vllm/model_executor/models/pixtral.py b/vllm/model_executor/models/pixtral.py index f7f46770057e2..d14b89d6b3f85 100644 --- a/vllm/model_executor/models/pixtral.py +++ b/vllm/model_executor/models/pixtral.py @@ -331,6 +331,7 @@ class VisionEncoderArgs: num_attention_heads: int rope_theta: float # for rope-2D image_token_id: int + adapter_bias: bool = True def _reshape_for_broadcast(freqs_cis: torch.Tensor, @@ -595,10 +596,10 @@ class VisionLanguageAdapter(nn.Module): self.w_in = nn.Linear( args.hidden_size, dim, - bias=True, + bias=args.adapter_bias, ) self.gelu = nn.GELU() - self.w_out = nn.Linear(dim, dim, bias=True) + self.w_out = nn.Linear(dim, dim, bias=args.adapter_bias) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.w_out(self.gelu(self.w_in(x)))