[Model] Systematic support for fp32 head, pooling models part (#23810)

Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-09-09 07:29:50 -07:00
committed by GitHub
parent a55cf41a09
commit 19332c0479
14 changed files with 166 additions and 61 deletions
+8 -8
View File
@@ -637,14 +637,14 @@ class GteNewForSequenceClassification(nn.Module, SupportsCrossEncoding):
self.new = GteNewModel(vllm_config=vllm_config,
prefix=prefix,
add_pooling_layer=True)
self.classifier = RowParallelLinear(config.hidden_size,
config.num_labels,
input_is_parallel=False,
bias=True,
quant_config=quant_config,
prefix=maybe_prefix(
prefix, "classifier"),
return_bias=False)
self.classifier = ReplicatedLinear(
config.hidden_size,
config.num_labels,
bias=True,
quant_config=quant_config,
params_dtype=vllm_config.model_config.head_dtype,
prefix=maybe_prefix(prefix, "classifier"),
return_bias=False)
pooler_config = vllm_config.model_config.pooler_config
assert pooler_config is not None