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

[feat] Support extra_buffer in Mamba2-based models (#15829)

Browse Source 
Signed-off-by: Roi Koren <roik@nvidia.com>
This commit is contained in:
roikoren755
2026-05-26 16:03:29 +08:00
committed by GitHub
parent 7e6e5efe51
commit e958f4561f
17 changed files with 405 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/models/test_granite_moe_hybrid.py
+33
View File
@@ -0,0 +1,33 @@
import unittest
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
GRANITE_MOE_HYBRID_MODEL = "ibm-granite/granite-4.0-h-micro"
class TestGraniteMoeHybrid(GSM8KMixin, DefaultServerBase):
model = GRANITE_MOE_HYBRID_MODEL
gsm8k_accuracy_thres = 0.78
class TestGraniteMoeHybridExtraBuffer(
GSM8KMixin, KLDivergenceMixin, PrefixCacheBranchingMixin, DefaultServerBase
):
model = GRANITE_MOE_HYBRID_MODEL
cache_chunk_size = 256
gsm8k_accuracy_thres = 0.78
kl_div_thres = 0.002
kl_div_thres_prefill = 0.02
other_args = [
"--mem-fraction-static",
"0.8",
"--mamba-scheduler-strategy",
"extra_buffer",
]
if __name__ == "__main__":
unittest.main()
test/manual/models/test_nvidia_nemotron_nano_v2.py
+65 -52
View File
@@ -2,91 +2,104 @@ import unittest
from sglang.srt.utils import is_blackwell
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
NVIDIA_NEMOTRON_NANO_V2_MODEL = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
model = NVIDIA_NEMOTRON_NANO_V2_MODEL
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256"]
class TestNvidiaNemotronNanoV2BF16PP(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
model = NVIDIA_NEMOTRON_NANO_V2_MODEL
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256", "--pp-size", "2"]
class TestNvidiaNemotronNanoV2FP8(GSM8KMixin, DefaultServerBase):
class TestNvidiaNemotronNanoV2BF16ExtraBuffer(
GSM8KMixin, KLDivergenceMixin, PrefixCacheBranchingMixin, DefaultServerBase
):
model = NVIDIA_NEMOTRON_NANO_V2_MODEL
cache_chunk_size = 256
gsm8k_accuracy_thres = 0.87
kl_div_thres = 0.002
kl_div_thres_prefill = 0.01
other_args = [
"--max-mamba-cache-size",
"256",
"--mem-fraction-static",
"0.8",
"--mamba-scheduler-strategy",
"extra_buffer",
]
class TestNvidiaNemotronNanoV2FP8(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8"
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256"]
@unittest.skipIf(not is_blackwell(), "NVFP4 only supported on blackwell")
class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.855
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-NVFP4"
gsm8k_accuracy_thres = 0.855
other_args = ["--max-mamba-cache-size", "256"]
@unittest.skip(
"STANDALONE speculative decoding does not yet support target and draft models "
"with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
)
SPECULATIVE_DECODING_OTHER_ARGS = [
"--speculative-algorithm",
"STANDALONE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"5",
"--speculative-draft-model-path",
"meta-llama/Llama-3.2-1B",
"--speculative-draft-load-format",
"dummy",
"--max-running-requests",
"8",
"--max-total-tokens",
"2048",
"--json-model-override-args",
'{"vocab_size": 131072, "hidden_size": 4480}',
]
class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase):
model = NVIDIA_NEMOTRON_NANO_V2_MODEL
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
"STANDALONE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"5",
"--speculative-draft-model-path",
"meta-llama/Llama-3.2-1B",
"--speculative-draft-load-format",
"dummy",
"--max-running-requests",
"8",
"--max-total-tokens",
"2048",
"--json-model-override-args",
'{"vocab_size": 131072}',
other_args = SPECULATIVE_DECODING_OTHER_ARGS + [
"--disable-radix-cache",
]
class TestNvidiaNemotronNanoV2SpeculativeDecodingExtraBuffer(
GSM8KMixin, DefaultServerBase
):
model = NVIDIA_NEMOTRON_NANO_V2_MODEL
gsm8k_accuracy_thres = 0.87
other_args = SPECULATIVE_DECODING_OTHER_ARGS + [
"--mamba-scheduler-strategy",
"extra_buffer",
]
@unittest.skip(
"STANDALONE speculative decoding does not yet support target and draft models "
"with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
)
class TestNvidiaNemotronNanoV2SpeculativeDecodingBF16Cache(
GSM8KMixin, DefaultServerBase
):
model = NVIDIA_NEMOTRON_NANO_V2_MODEL
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
"STANDALONE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"5",
"--speculative-draft-model-path",
"meta-llama/Llama-3.2-1B",
"--speculative-draft-load-format",
"dummy",
"--max-running-requests",
"8",
"--max-total-tokens",
"2048",
"--json-model-override-args",
'{"vocab_size": 131072}',
other_args = SPECULATIVE_DECODING_OTHER_ARGS + [
"--disable-radix-cache",
"--mamba-ssm-dtype",
"bfloat16",
]