minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[misc] Use --cuda-graph-max-bs-decode in tests, examples, and docs (#29591)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-06-28 18:38:28 -07:00
committed by GitHub
parent 3217410cf6
commit 909123ddb8
140 changed files with 304 additions and 290 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/dsv4/test_b200_pro.py
+3 -3
View File
@@ -60,7 +60,7 @@ class TestB200ProBalanced(DSV4ProAime25TestBase):
"2",
"--mem-fraction-static",
"0.82",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"64",
"--max-running-requests",
"128",
@@ -83,7 +83,7 @@ class TestB200ProMaxThroughput(DSV4ProAime25TestBase):
"deepep",
"--mem-fraction-static",
"0.82",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"64",
"--max-running-requests",
"256",
@@ -108,7 +108,7 @@ class TestB200ProCP(DSV4ProAime25TestBase):
"16384",
"--mem-fraction-static",
"0.78",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"256",
"--max-running-requests",
"256",
test/manual/dsv4/test_b300_pro.py
+3 -3
View File
@@ -62,7 +62,7 @@ class TestB300ProBalanced(DSV4ProAime25TestBase):
"2",
"--mem-fraction-static",
"0.82",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"64",
"--max-running-requests",
"128",
@@ -85,7 +85,7 @@ class TestB300ProMaxThroughput(DSV4ProAime25TestBase):
"deepep",
"--mem-fraction-static",
"0.82",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"64",
"--max-running-requests",
"256",
@@ -110,7 +110,7 @@ class TestB300ProCP(DSV4ProAime25TestBase):
"16384",
"--mem-fraction-static",
"0.78",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"256",
"--max-running-requests",
"256",
test/manual/dsv4/test_dsv4_flash_mtp_dp4.py
+2 -2
View File
@@ -37,7 +37,7 @@ DSV4_FLASH_ENV = {
"SGLANG_DSV4_FP4_EXPERTS": "0",
# MTP runs ~num_draft_tokens forward passes per step, so the deepep
# dispatch input size scales by that factor. Default 256 (used by the
# plain server) overflows once cuda-graph-max-bs * num_draft_tokens
# plain server) overflows once cuda-graph-max-bs-decode * num_draft_tokens
# > 256. 1024 covers bs=128 * 4 draft tokens with headroom.
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "1024",
}
@@ -66,7 +66,7 @@ class DSV4FlashMTPServerBase(CustomTestCase):
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
test/manual/dsv4/test_dsv4_flash_sanity_dp4.py
+3 -3
View File
@@ -65,7 +65,7 @@ class TestDSV4FlashTP4DP4(
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
@@ -101,7 +101,7 @@ class TestDSV4FlashTP4EP(
# No --enable-dp-attention by design: covers TP-attn path.
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"64",
@@ -139,7 +139,7 @@ class TestDSV4FlashTP4DP4ChunkedPrefillLarge(
"deepep",
"--chunked-prefill-size",
"16384",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
test/manual/dsv4/test_dsv4_pd_disagg_nixl.py
+2 -2
View File
@@ -61,7 +61,7 @@ class TestDSV4FlashPDDisaggNIXL(PDDisaggregationServerBase):
"deepep",
"--deepep-config",
DEEPEP_CONFIG,
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
@@ -103,7 +103,7 @@ class TestDSV4FlashPDDisaggNIXL(PDDisaggregationServerBase):
"deepep",
"--deepep-config",
DEEPEP_CONFIG,
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
test/manual/dsv4/test_dsv4_swa_radix_retract.py
+1 -1
View File
@@ -66,7 +66,7 @@ class TestDSV4FlashSWARadixRetract(CustomTestCase):
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
test/manual/dsv4/test_gb300_pro.py
+3 -3
View File
@@ -62,7 +62,7 @@ class TestGB300ProBalanced(DSV4ProAime25TestBase):
"2",
"--mem-fraction-static",
"0.9",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
@@ -85,7 +85,7 @@ class TestGB300ProMaxThroughput(DSV4ProAime25TestBase):
"deepep",
"--mem-fraction-static",
"0.9",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
@@ -110,7 +110,7 @@ class TestGB300ProCP(DSV4ProAime25TestBase):
"16384",
"--mem-fraction-static",
"0.88",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"256",
"--max-running-requests",
"256",
test/manual/dsv4/test_h200_fp8_flash.py
+2 -2
View File
@@ -53,7 +53,7 @@ class TestH200Fp8FlashBalanced(DSV4FlashAime25TestBase):
"1",
"--speculative-num-draft-tokens",
"2",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"128",
@@ -77,7 +77,7 @@ class TestH200Fp8FlashMaxThroughput(DSV4FlashAime25TestBase):
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",
test/manual/dsv4/test_h200_fp8_pro.py
+3 -3
View File
@@ -44,7 +44,7 @@ class TestH200Fp8ProLowLatency(DSV4ProAime25TestBase):
*multinode_args(2),
"--moe-a2a-backend",
"deepep",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"8",
"--max-running-requests",
"32",
@@ -88,7 +88,7 @@ class TestH200Fp8ProBalanced(DSV4ProAime25TestBase):
"2",
"--mem-fraction-static",
"0.88",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"8",
"--max-running-requests",
"32",
@@ -114,7 +114,7 @@ class TestH200Fp8ProMaxThroughput(DSV4ProAime25TestBase):
"deepep",
"--mem-fraction-static",
"0.88",
"--cuda-graph-max-bs",
"--cuda-graph-max-bs-decode",
"128",
"--max-running-requests",
"256",