minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[5/n] Lora support cuda graph (#21647)

Browse Source
This commit is contained in:
Ethan (Yusheng) Su
2026-04-04 00:31:46 -07:00
committed by GitHub
parent a94c3804c2
commit ff8e47edf9
15 changed files with 311 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/lora/test_lora_gpt_oss_20b_logprob_diff.py
-2
View File
@@ -44,7 +44,6 @@ LORA_HF_REPO = "yushengsu/lora-diff-gpt-oss-20b"
LORA_BACKEND = "triton"
MAX_LORA_RANK = 32
TP_SIZE = 4
DISABLE_CUDA_GRAPH = True
MOE_RUNNER_BACKEND = "triton"
EXPERTS_SHARED_OUTER_LORAS = True
PREFILL_ATTENTION_BACKEND = "fa4"
@@ -86,7 +85,6 @@ class TestLoRAGptOss20BLogprobDiff(CustomTestCase):
lora_paths={"my_lora": adapter_path},
lora_backend=LORA_BACKEND,
attention_backend="flashinfer",
disable_cuda_graph=DISABLE_CUDA_GRAPH,
moe_runner_backend=MOE_RUNNER_BACKEND,
experts_shared_outer_loras=EXPERTS_SHARED_OUTER_LORAS,
prefill_attention_backend=PREFILL_ATTENTION_BACKEND,
test/registered/lora/test_lora_qwen3_30b_a3b_instruct_2507_logprob_diff.py
-2
View File
@@ -44,7 +44,6 @@ LORA_HF_REPO = "yushengsu/lora-diff-Qwen3-30B-A3B-Instruct-2507"
LORA_BACKEND = "triton"
MAX_LORA_RANK = 32
TP_SIZE = 4
DISABLE_CUDA_GRAPH = True
MOE_RUNNER_BACKEND = "triton"
EXPERTS_SHARED_OUTER_LORAS = True
PREFILL_ATTENTION_BACKEND = "fa4"
@@ -86,7 +85,6 @@ class TestLoRAQwen3_30B_A3B_Instruct_2507_LogprobDiff(CustomTestCase):
lora_paths={"my_lora": adapter_path},
lora_backend=LORA_BACKEND,
attention_backend="flashinfer",
disable_cuda_graph=DISABLE_CUDA_GRAPH,
moe_runner_backend=MOE_RUNNER_BACKEND,
experts_shared_outer_loras=EXPERTS_SHARED_OUTER_LORAS,
prefill_attention_backend=PREFILL_ATTENTION_BACKEND,
test/registered/lora/test_lora_qwen3_8b_logprob_diff.py
-2
View File
@@ -47,7 +47,6 @@ LORA_HF_REPO = "yushengsu/lora-diff-Qwen3-8B"
LORA_BACKEND = "triton"
MAX_LORA_RANK = 32
TP_SIZE = 1
DISABLE_CUDA_GRAPH = True
PREFILL_ATTENTION_BACKEND = "fa4"
DECODE_ATTENTION_BACKEND = "fa4"
@@ -139,7 +138,6 @@ class TestLoRAQwen3_8BLogprobDiff(CustomTestCase):
lora_paths={"my_lora": adapter_path},
lora_backend=LORA_BACKEND,
attention_backend="flashinfer",
disable_cuda_graph=DISABLE_CUDA_GRAPH,
prefill_attention_backend=PREFILL_ATTENTION_BACKEND,
decode_attention_backend=DECODE_ATTENTION_BACKEND,
)
test/registered/lora/test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff.py
-2
View File
@@ -44,7 +44,6 @@ LORA_HF_REPO = "yushengsu/lora-diff-Qwen3-VL-30B-A3B-Instruct"
LORA_BACKEND = "triton"
MAX_LORA_RANK = 32
TP_SIZE = 4
DISABLE_CUDA_GRAPH = True
MOE_RUNNER_BACKEND = "triton"
EXPERTS_SHARED_OUTER_LORAS = True
PREFILL_ATTENTION_BACKEND = "fa4"
@@ -86,7 +85,6 @@ class TestLoRAQwen3VL_30B_A3B_Instruct_LogprobDiff(CustomTestCase):
lora_paths={"my_lora": adapter_path},
lora_backend=LORA_BACKEND,
attention_backend="flashinfer",
disable_cuda_graph=DISABLE_CUDA_GRAPH,
moe_runner_backend=MOE_RUNNER_BACKEND,
experts_shared_outer_loras=EXPERTS_SHARED_OUTER_LORAS,
prefill_attention_backend=PREFILL_ATTENTION_BACKEND,