Update rope_scaling to rope_parameters in preparation for Transformers v5 (#28542)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-11-19 09:06:36 -08:00
committed by GitHub
parent d44e9df7d4
commit a8b70304d6
104 changed files with 542 additions and 910 deletions
+12 -9
View File
@@ -155,25 +155,28 @@ class Gemma3Attention(nn.Module):
self.k_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
layer_idx = extract_layer_index(prefix)
self.is_sliding = config.layer_types[layer_idx] == "sliding_attention"
layer_type = config.layer_types[layer_idx]
self.is_sliding = layer_type == "sliding_attention"
sliding_window = config.sliding_window if self.is_sliding else None
# Initialize the rotary embedding.
if self.is_sliding:
# Local attention. Override the values in config.json.
self.rope_theta = config.rope_local_base_freq
self.rope_scaling = {"rope_type": "default"}
if layer_type in config.rope_parameters:
# Transformers v5 rope config.
rope_parameters = config.rope_parameters[layer_type]
else:
# Transformers v4 rope config.
# Global attention. Use the values in config.json.
self.rope_theta = config.rope_theta
self.rope_scaling = config.rope_scaling
rope_parameters = config.rope_parameters.copy()
# Local attention. Override the values in config.json.
if self.is_sliding:
rope_parameters["rope_theta"] = config.rope_local_base_freq
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
base=self.rope_theta,
rope_parameters=rope_parameters,
is_neox_style=True,
rope_scaling=self.rope_scaling,
)
if getattr(config, "is_causal", True):