@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 1838cd4860

Fork 0

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 14, 05:30 UTC

Revert "Add batch invariant kernel override for FlashInfer backend [2/n]" (#26220)

Cyrus Leung committed Oct 4, 2025 at 17:45 UTC 1838cd4860cf96972812c80a1cfc788f9ba3584b

3 files changed +29 -84

tests/v1/generation/test_batch_invariance.py

+23 -40

@@ -76,21 +76,18 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle():

76 seed.

77 - Keep max_tokens and max_model_len bounded for speed and memory use.

78 """

79 - seed = int(os.getenv("VLLM_TEST_SEED", "12345"))

80 - random.seed(seed)

79 + random.seed(12345)

81 # Allow overrides from environment (useful for CI tuning)

82 # "facebook/opt-125m" is too small, doesn't reliably test determinism

83 model = os.getenv("VLLM_TEST_MODEL", "Qwen/Qwen3-1.7B")

84 num_trials = int(os.getenv("VLLM_NEEDLE_TRIALS", "5"))

86 - max_batch_size = int(os.getenv("VLLM_NEEDLE_BATCH_SIZE", "128"))

87 - min_random_prompt = int(os.getenv("VLLM_MIN_PROMPT", "1024"))

88 - max_random_prompt = int(os.getenv("VLLM_MAX_PROMPT", "2048"))

89 - assert max_batch_size >= 2, "Batch size should be >= 2 to mix needle."

85 + batch_size = int(os.getenv("VLLM_NEEDLE_BATCH_SIZE", "64"))

86 + assert batch_size >= 2, "Batch size should be >= 2 to mix needle."

88 # Keep GPU memory usage low to avoid startup allocation failures.

92 - gpu_mem_util = float(os.getenv("VLLM_GPU_MEMORY_UTILIZATION", "0.4"))

93 - max_model_len = int(os.getenv("VLLM_MAX_MODEL_LEN", "5120"))

89 + gpu_mem_util = float(os.getenv("VLLM_GPU_MEMORY_UTILIZATION", "0.3"))

90 + max_model_len = int(os.getenv("VLLM_MAX_MODEL_LEN", "4096"))

91 swap_space_gb = int(os.getenv("VLLM_SWAP_SPACE_GB", "4"))

93 # Sampling parameters: longer outputs with a more random-sounding

@@ -114,7 +111,7 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle():

111 # Engine with bs=1 behavior

112 llm_bs1 = LLM_with_max_seqs(

113 model=model,

117 - max_num_seqs=128,

114 + max_num_seqs=1,

115 gpu_memory_utilization=gpu_mem_util,

116 max_model_len=max_model_len,

117 swap_space=swap_space_gb,

@@ -129,7 +126,7 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle():

126 # Engine with larger batch limit (e.g., 64)

127 llm_bsN = LLM_with_max_seqs(

128 model=model,

132 - max_num_seqs=128,

129 + max_num_seqs=batch_size,

130 gpu_memory_utilization=gpu_mem_util,

131 max_model_len=max_model_len,

132 swap_space=swap_space_gb,

@@ -138,17 +135,15 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle():

135 mismatches = 0

136

137 for trial in range(num_trials):

141 - # Create a batch of size `max_batch_size` and insert the needle at

138 + # Create a batch of size `batch_size` and insert the needle at

139 # a random index

140 prompts: list[str] = []

144 - batch_size = random.randint(max_batch_size // 2, max_batch_size)

141 needle_pos = random.randint(0, batch_size - 1)

142 for i in range(batch_size):

143 if i == needle_pos:

144 prompts.append(needle_prompt)

145 else:

150 - prompts.append(

151 - _random_prompt(min_random_prompt, max_random_prompt))

146 + prompts.append(_random_prompt())

147

148 # Generate with the larger-batch engine

149 outputs = llm_bsN.generate(prompts, sampling)

@@ -159,19 +154,17 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle():

154 text = needle_output.outputs[0].text

155

156 if text != baseline_text:

162 - print(

163 - f"{text}\n\n== Not the same as ==\n\n{baseline_text}\n\n")

157 mismatches += 1

158

159 passes = num_trials - mismatches

160 # Dump how many passed vs failed

161 print(f"[determinism] total={num_trials}, passed={passes}, "

169 - f"failed={mismatches}, max_batch_size={max_batch_size}")

162 + f"failed={mismatches}, batch_size={batch_size}")

163

164 if mismatches > 0:

165 pytest.fail(

166 f"Nondeterministic outputs detected: {mismatches} failed out "

174 - f"of {num_trials} trials (max_batch_size={max_batch_size}).")

167 + f"of {num_trials} trials (batch_size={batch_size}).")

168

169 finally:

170 # Ensure engines are shutdown to free GPU/VRAM across test sessions

@@ -203,14 +196,9 @@ def _extract_step_logprobs(request_output):

196 not torch.cuda.is_available(),

197 reason="Requires CUDA to match production inference path.",

198 )

206 -@pytest.mark.parametrize("backend", ["FLEX_ATTENTION", "FLASHINFER"])

207 -def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN(backend):

199 +def test_logprobs_bitwise_batch_invariance_bs1_vs_bs2():

200

209 - backend = os.getenv("VLLM_ATTENTION_BACKEND", backend)

210 - os.environ["VLLM_ATTENTION_BACKEND"] = backend

211 -

212 - seed = int(os.getenv("VLLM_TEST_SEED", "12345"))

213 - random.seed(seed)

201 + #model_name = os.getenv("VLLM_TEST_MODEL", "facebook/opt-125m")

202 model_name = os.getenv("VLLM_TEST_MODEL", "Qwen/Qwen3-1.7B")

203 tp_size = int(os.getenv("VLLM_TEST_TP_SIZE", "1"))

204

@@ -224,15 +212,10 @@ def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN(backend):

212 prompts = [

213 "The capital of France is",

214 "The capital of Germany is",

227 - _random_prompt(10, 1024),

228 - _random_prompt(10, 1024),

229 - _random_prompt(10, 1024),

230 - _random_prompt(10, 1024),

231 - _random_prompt(10, 1024),

215 ]

216

217 sp = SamplingParams(

235 - temperature=0.6,

218 + temperature=0.0,

219 top_p=1.0,

220 max_tokens=8,

221 # Seed shouldn't matter at temperature=0, but keeping it stable anyway.

@@ -251,25 +234,25 @@ def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN(backend):

234 "enable logprobs return to run this test.")

235 bs1_logprobs_per_prompt.append(step_logprobs)

236

254 - # BS=N: run prompts in a batch and collect logprobs per step for each

237 + # BS=2: run prompts in a batch and collect logprobs per step for each

238 # prompt.

239 outs_batched = llm.generate(prompts, sp, use_tqdm=False)

240 assert len(outs_batched) == len(prompts)

258 - bsN_logprobs_per_prompt = []

241 + bs2_logprobs_per_prompt = []

242 for o in outs_batched:

243 step_logprobs = _extract_step_logprobs(o)

244 if step_logprobs is None:

245 pytest.skip("Logits are not available on RequestOutput; "

246 "enable logprobs return to run this test.")

264 - bsN_logprobs_per_prompt.append(step_logprobs)

247 + bs2_logprobs_per_prompt.append(step_logprobs)

248

266 - # Compare step-by-step logprobs for each prompt between BS=1 and BS=N runs.

267 - for i, (logprobs_bs1, logprobs_bsN) in enumerate(

268 - zip(bs1_logprobs_per_prompt, bsN_logprobs_per_prompt)):

269 - assert len(logprobs_bs1) == len(logprobs_bsN), (

249 + # Compare step-by-step logprobs for each prompt between BS=1 and BS=2 runs.

250 + for i, (logprobs_bs1, logprobs_bs2) in enumerate(

251 + zip(bs1_logprobs_per_prompt, bs2_logprobs_per_prompt)):

252 + assert len(logprobs_bs1) == len(logprobs_bs2), (

253 f"Different number of generation steps for prompt index {i}: "

271 - f"{len(logprobs_bs1)} (BS=1) vs {len(logprobs_bsN)} (BS=N)")

272 - for t, (a, b) in enumerate(zip(logprobs_bs1, logprobs_bsN)):

254 + f"{len(logprobs_bs1)} (BS=1) vs {len(logprobs_bs2)} (BS=2)")

255 + for t, (a, b) in enumerate(zip(logprobs_bs1, logprobs_bs2)):

256 assert a.shape == b.shape, (

257 f"Logits shape mismatch at prompt {i}, step {t}: "

258 f"{a.shape} vs {b.shape}")

vllm/model_executor/layers/batch_invariant.py

+1 -12

@@ -8,12 +8,8 @@ from typing import Any, Union

9 import torch

11 -import vllm.envs as envs

12 -from vllm.logger import init_logger

11 from vllm.triton_utils import tl, triton

15 -logger = init_logger(__name__)

16 -

14 def _matmul_launch_metadata(grid: Callable[..., Any], kernel: Any,

15 args: dict[str, Any]) -> dict[str, Any]:

@@ -561,12 +557,5 @@ def vllm_kernel_override_batch_invariant():

557 def init_batch_invariance():

558 # this will hit all the csrc overrides as well

559 if vllm_kernel_override_batch_invariant():

564 - curr_attn_backend = envs.VLLM_ATTENTION_BACKEND

565 - supported_backends = ["FLEX_ATTENTION", "FLASHINFER"]

566 - if curr_attn_backend not in supported_backends:

567 - warning = "Forcibly updating attention backend to" \

568 - f" {supported_backends[0]} for batch_invariant. " \

569 - f" Supported backends: {supported_backends}."

570 - logger.warning_once(warning)

571 - os.environ["VLLM_ATTENTION_BACKEND"] = supported_backends[0]

560 + os.environ["VLLM_ATTENTION_BACKEND"] = "FLEX_ATTENTION"

561 enable_batch_invariant_mode()

vllm/v1/attention/backends/flashinfer.py

+5 -32

@@ -20,8 +20,6 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,

20 AttentionType)

21 from vllm.config import CUDAGraphMode, VllmConfig

22 from vllm.logger import init_logger

23 -from vllm.model_executor.layers.batch_invariant import (

24 - vllm_kernel_override_batch_invariant)

23 from vllm.model_executor.layers.quantization.utils.quant_utils import (

24 QuantKey, kFp8StaticTensorSym, kNvfp4Quant)

25 from vllm.platforms import current_platform

@@ -44,7 +42,6 @@ from vllm.v1.attention.backends.utils import (AttentionCGSupport,

42 from vllm.v1.kv_cache_interface import AttentionSpec

44 FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024

47 -FLASHINFER_WORKSPACE_BUFFER_SIZE_BATCH_INVARIANT = 2048 * 1024 * 1024

46 FP8_DTYPE = current_platform.fp8_dtype()

47 FP4_DTYPE = torch.uint8

@@ -266,15 +263,6 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):

263 self._prefill_wrapper = None # Wrapper for prefill/append

264 self._decode_wrapper = None # Wrapper for decode (general shape)

265

269 - if vllm_kernel_override_batch_invariant():

270 - self.decode_fixed_split_size = 2048

271 - self.prefill_fixed_split_size = 4096

272 - self.disable_split_kv = True

273 - else:

274 - self.decode_fixed_split_size = -1

275 - self.prefill_fixed_split_size = -1

276 - self.disable_split_kv = False

277 -

266 self.compilation_config = vllm_config.compilation_config

267 max_num_pages_per_req = cdiv(self.model_config.max_model_len,

268 self.kv_cache_spec.block_size)

@@ -368,12 +356,10 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):

356

357 def _get_workspace_buffer(self):

358 if self._workspace_buffer is None:

371 - buffer_size = FLASHINFER_WORKSPACE_BUFFER_SIZE

372 - if vllm_kernel_override_batch_invariant():

373 - buffer_size = FLASHINFER_WORKSPACE_BUFFER_SIZE_BATCH_INVARIANT

374 - self._workspace_buffer = torch.zeros(buffer_size,

375 - dtype=torch.uint8,

376 - device=self.device)

359 + self._workspace_buffer = torch.zeros(

360 + FLASHINFER_WORKSPACE_BUFFER_SIZE,

361 + dtype=torch.uint8,

362 + device=self.device)

363 return self._workspace_buffer

364

365 def _get_prefill_wrapper(self):

@@ -629,8 +615,6 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):

615 logits_soft_cap=self.logits_soft_cap,

616 q_data_type=self.q_data_type,

617 kv_data_type=self.kv_cache_dtype,

632 - fixed_split_size=self.prefill_fixed_split_size,

633 - disable_split_kv=self.disable_split_kv,

618 )

619 else:

620 attn_metadata.qo_indptr_gpu = qo_indptr_cpu.to(

@@ -684,8 +668,6 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):

668 logits_soft_cap=self.logits_soft_cap,

669 q_data_type=self.q_data_type,

670 kv_data_type=self.kv_cache_dtype,

687 - fixed_split_size=self.decode_fixed_split_size,

688 - disable_split_kv=self.disable_split_kv,

671 )

672 return attn_metadata

673

@@ -1066,8 +1048,6 @@ def fast_plan_decode(

1048 rope_scale: Optional[float] = None,

1049 rope_theta: Optional[float] = None,

1050 non_blocking: bool = True,

1069 - fixed_split_size: int = -1,

1070 - disable_split_kv: bool = False,

1051 ) -> None:

1052 """

1053 A faster version of BatchDecodeWithPagedKVCacheWrapper::plan used for

@@ -1105,10 +1085,6 @@ def fast_plan_decode(

1085 rope_scale,

1086 rope_theta,

1087 non_blocking,

1108 - None, # block_tables

1109 - None, # seq_lens

1110 - fixed_split_size,

1111 - disable_split_kv,

1088 )

1089 self.vllm_first_call = False

1090 return

@@ -1154,7 +1130,7 @@ def fast_plan_decode(

1130 qo_indptr_host = _get_range_buf(batch_size + 1, "cpu")

1131

1132 try:

1157 - # Make sure we pass exactly 18 arguments for tensor core version

1133 + # Make sure we pass exactly 15 arguments for tensor core version

1134 self._plan_info = self._cached_module.plan(

1135 self._float_workspace_buffer,

1136 self._int_workspace_buffer,

@@ -1171,9 +1147,6 @@ def fast_plan_decode(

1147 head_dim,

1148 head_dim,

1149 False, # causal

1174 - window_left,

1175 - fixed_split_size,

1176 - disable_split_kv,

1150 )

1151 except Exception as e:

1152 raise RuntimeError(f"Error in tensor core plan: {e}") from e