[Bugfix] LoRA for DeepSeek V3.2 (#35077)

Signed-off-by: Hollow Man <hollowman@opensuse.org>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟
2026-04-22 19:33:50 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Jee Jee Li
parent 04eac6ba24
commit a250f1bd5f
13 changed files with 614 additions and 37 deletions
@@ -214,6 +214,19 @@ def select_unquantized_moe_backend(
return backend, k_cls
raise ValueError(_make_log_unsupported(backend, reason))
# LoRA needs Triton's unfused activation/reduction hooks. Selecting the
# backend here ensures weights stay in a LoRA-compatible layout instead of
# being permuted for a backend like FlashInfer or AITER during load.
if moe_config.is_lora_enabled:
backend = UnquantizedMoeBackend.TRITON
if activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
backend = UnquantizedMoeBackend.BATCHED_TRITON
return _return_or_raise(
backend,
moe_config,
activation_format,
)
runner_backend = moe_config.moe_backend
if runner_backend != "auto":
requested_backend = map_unquantized_backend(runner_backend)
@@ -356,6 +356,12 @@ def get_and_maybe_dequant_weights(
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
# LoRA linear wrappers store quantization metadata on `base_layer`.
# Unwrap here so callers can pass either a raw linear layer or its LoRA
# wrapper without special-casing.
while hasattr(layer, "base_layer") and hasattr(layer.base_layer, "quant_method"):
layer = layer.base_layer
weight = get_attribute_fallback(layer, ["weight", "qweight", "weight_packed"])
# Unquantized layer: just return base weights