A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 19:00 UTC
[LoRA] LoRA PDL improvement (#31660)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Jan 5, 2026 at 16:28 UTC
d5503ca7f93ee756604c8dac0adeec9526c40802
1 file changed
+4
-2
vllm/lora/ops/triton_ops/fused_moe_lora_op.py
+4
-2
@@ -163,15 +163,17 @@ def _fused_moe_lora_kernel(
163
# accumulator
164
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
165
166
- # GDC wait waits for ALL programs in the prior kernel to complete
167
- # before continuing.
166
if USE_GDC and not IS_PRIMARY:
167
tl.extra.cuda.gdc_wait()
168
169
for k in range(0, grid_k):
170
k_remaining = K - k * (BLOCK_SIZE_K * SPLIT_K)
171
+ # GDC wait waits for ALL programs in the prior kernel to complete
172
+ # before continuing.
173
# pre-fetch lora weight
174
b = tl.load(b_ptrs, mask=offs_k[:, None] < k_remaining, other=0.0)
175
+ if USE_GDC and not IS_PRIMARY:
176
+ tl.extra.cuda.gdc_wait()
177
a = tl.load(
178
a_ptrs,
179
mask=token_mask[:, None] & (offs_k[None, :] < k_remaining),