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

[Qwen3.5] Enable MTP spec_v2 and add test for nvidia/Qwen3.5-397B-A17B-NVFP4 (#19391)

Browse Source
This commit is contained in:
hlu1
2026-03-04 14:01:25 -08:00
committed by GitHub
parent 0ee9d3c8e9
commit 9457c049e1
8 changed files with 252 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/disaggregation/decode.py
-5
View File
@@ -183,11 +183,6 @@ class HybridMambaDecodeReqToTokenPool(HybridReqToTokenPool):
pre_alloc_size=pre_alloc_size,
)
if envs.SGLANG_ENABLE_SPEC_V2.get() and not enable_mamba_extra_buffer:
raise ValueError(
"Spec v2 requires mamba scheduler strategy `extra_buffer` for mamba models. "
"Please set `--mamba-scheduler-strategy extra_buffer`."
)
self.mamba_ping_pong_track_buffer_size = 2 if enable_overlap_schedule else 1
self.enable_mamba_extra_buffer = enable_mamba_extra_buffer
self.enable_memory_saver = enable_memory_saver
python/sglang/srt/mem_cache/memory_pool.py
-5
View File
@@ -465,11 +465,6 @@ class HybridReqToTokenPool(ReqToTokenPool):
device=device,
enable_memory_saver=enable_memory_saver,
)
if envs.SGLANG_ENABLE_SPEC_V2.get() and not enable_mamba_extra_buffer:
raise ValueError(
"Spec v2 requires mamba scheduler strategy `extra_buffer` for mamba models. "
"Please set `--mamba-scheduler-strategy extra_buffer`."
)
self.mamba_ping_pong_track_buffer_size = 2 if enable_overlap_schedule else 1
self.enable_mamba_extra_buffer = enable_mamba_extra_buffer
python/sglang/srt/model_executor/forward_batch_info.py
+1 -1
View File
@@ -684,7 +684,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
mm_input, self.seq_lens_cpu[batch_idx]
)
mrope_positions_list[batch_idx] = mrope_positions
elif self.forward_mode.is_extend():
elif self.forward_mode.is_extend(include_draft_extend_v2=True):
extend_seq_len, extend_prefix_len = (
batch.extend_seq_lens[batch_idx],
batch.extend_prefix_lens[batch_idx],
python/sglang/srt/models/qwen3_5_mtp.py
+1 -1
View File
@@ -111,7 +111,7 @@ class Qwen3_5ForCausalLMMTP(nn.Module):
if (
forward_batch.forward_mode.is_extend()
and forward_batch.contains_mm_inputs()
and not forward_batch.forward_mode.is_draft_extend()
and not forward_batch.forward_mode.is_draft_extend(include_v2=True)
):
assert input_embeds is not None
input_embeds = torch.cat(
python/sglang/srt/server_args.py
+5 -4
View File
@@ -1896,10 +1896,11 @@ class ServerArgs:
self.disable_radix_cache = True
self.disable_overlap_schedule = False
else:
logger.warning(
f"Disabling radix cache since speculative decoding for {model_arch} is not supported with radix cache yet."
)
self.disable_radix_cache = True
if not self.disable_radix_cache:
raise ValueError(
f"Speculative decoding for {model_arch} is not compatible with radix cache when using --mamba-scheduler-strategy no_buffer."
"To use radix cache with speculative decoding, please use --mamba-scheduler-strategy extra_buffer and set SGLANG_ENABLE_SPEC_V2=1."
)
def _handle_sampling_backend(self):
if self.sampling_backend is None:
python/sglang/srt/speculative/eagle_worker_v2.py
+4
View File
@@ -465,6 +465,7 @@ class EagleDraftWorker(BaseDraftWorker):
batch: ModelWorkerBatch,
target_hidden_states: torch.Tensor,
next_token_ids: torch.Tensor,
mm_input_embeds: Optional[torch.Tensor] = None,
):
"""
Run draft model extend to correctly fill the KV cache.
@@ -498,6 +499,8 @@ class EagleDraftWorker(BaseDraftWorker):
# Run forward
forward_batch = ForwardBatch.init_new(batch, self.draft_runner)
if mm_input_embeds is not None:
forward_batch.mm_input_embeds = mm_input_embeds
logits_output = self.draft_runner.forward(forward_batch).logits_output
# Update spec_info for the next draft step
@@ -668,6 +671,7 @@ class EAGLEWorkerV2(BaseSpecWorker):
model_worker_batch,
batch_output.logits_output.hidden_states,
batch_output.next_token_ids,
batch_output.logits_output.mm_input_embeds,
)
)
return batch_output
test/registered/4-gpu-models/test_qwen35_models.py
+240
View File
@@ -0,0 +1,240 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
# This eval harness applies the chat_template, which is critical for qwen3.5
# to get good accuracy on gsm8k
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=1000, suite="stage-c-test-4-gpu-b200")
QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
ACC_THRESHOLDS = {
QWEN35_FP4_MODEL: {"gsm8k": 0.95},
}
class TestQwen35FP4(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN35_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
"--mamba-ssm-dtype",
"bfloat16",
"--max-running-requests",
"128",
"--reasoning-parser",
"qwen3",
"--attention-backend",
"trtllm_mha",
"--quantization",
"modelopt_fp4",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
model=self.model,
eval_name="gsm8k",
num_shots=5,
num_examples=200,
max_tokens=16000,
num_threads=128,
repeat=1,
temperature=0.6,
top_p=0.95,
top_k=20,
base_url=self.base_url,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(metrics["score"], ACC_THRESHOLDS[self.model]["gsm8k"])
class TestQwen35FP4MTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN35_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
"--mamba-ssm-dtype",
"bfloat16",
"--max-running-requests",
"128",
"--reasoning-parser",
"qwen3",
"--attention-backend",
"trtllm_mha",
"--quantization",
"modelopt_fp4",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
model=self.model,
eval_name="gsm8k",
num_shots=5,
num_examples=200,
max_tokens=16000,
num_threads=128,
repeat=1,
temperature=0.6,
top_p=0.95,
top_k=20,
base_url=self.base_url,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(metrics["score"], ACC_THRESHOLDS[self.model]["gsm8k"])
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 3.3)
class TestQwen35FP4MTPV2(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN35_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
envs.SGLANG_ENABLE_SPEC_V2.set(True)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
"--mamba-ssm-dtype",
"bfloat16",
"--max-running-requests",
"128",
"--reasoning-parser",
"qwen3",
"--attention-backend",
"trtllm_mha",
"--quantization",
"modelopt_fp4",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
],
)
@classmethod
def tearDownClass(cls):
envs.SGLANG_ENABLE_SPEC_V2.set(False)
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
model=self.model,
eval_name="gsm8k",
num_shots=5,
num_examples=200,
max_tokens=16000,
num_threads=128,
repeat=1,
temperature=0.6,
top_p=0.95,
top_k=20,
base_url=self.base_url,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(metrics["score"], ACC_THRESHOLDS[self.model]["gsm8k"])
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 3.3)
if __name__ == "__main__":
unittest.main()
test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
+1
View File
@@ -72,6 +72,7 @@ class TestQwen3NextMTP(CustomTestCase):
"2048",
"--mamba-scheduler-strategy",
"no_buffer",
"--disable-radix-cache",
],
)