vllm.v1.worker.gpu.warmup
¶
Functions:
-
run_context_prefill_warmup–Run a prefill step over about
num_context_tokensof cached context. -
run_kv_cache_sizing_warmup–Run the steps an extensible KV cache measures its transient reserve on.
-
run_mixed_prefill_decode_warmup–Run a V2 mixed prefill+decode step through normal scheduler inputs,
-
warmup_kernels–Run scheduler-realistic prefill and decode steps to JIT compile kernels.
_cap_num_warmup_reqs(model_runner, num_reqs, num_tokens_per_req)
¶
Limit synthetic requests to the scheduler's token budget.
Source code in vllm/v1/worker/gpu/warmup.py
_fixed_draft_counts(model_runner)
¶
Verify fixed draft counts: adaptive costs are calibrated during capture, after the warmups.
Source code in vllm/v1/worker/gpu/warmup.py
_measure_sizing_step(model_runner)
¶
Record how far a sizing step grows torch's reservation from an emptied cache, as serving starts once the cache is sized. Unlike the allocated peak, this includes fragmentation, e.g. a large per-layer workspace that no longer fits the segments freed around it. Allocations the step keeps are not counted.
Source code in vllm/v1/worker/gpu/warmup.py
_reserved_block_count(num_tokens, group, *, vllm_config, num_lookahead_tokens, max_model_len, max_encoder_len, dcp_world_size=1)
¶
Number of blocks the scheduler would hold for a request of num_tokens.
Warmup hand-builds its SchedulerOutputs, so it must reserve what
KVCacheManager.allocate_slots reserves: the token range plus
num_lookahead_tokens, where the speculator writes the KV of its drafts.
A DCP-sharded block holds dcp_world_size times its tokens.
Source code in vllm/v1/worker/gpu/warmup.py
_run_step(worker_execute_model, worker_sample_tokens, scheduler_output)
¶
Run one synthetic step and wait for the GPU to finish it.
UvaBufferPool reuses a staging slot two writes later, which is safe in serving because the engine waits on each step's output. Warmup discards the outputs, so without this a slow GPU reads slots already overwritten.
Source code in vllm/v1/worker/gpu/warmup.py
_warmup_block_counter(model_runner)
¶
Bind _reserved_block_count to model_runner's reservation policy.
Source code in vllm/v1/worker/gpu/warmup.py
_warmup_token_ids(model_runner, num_tokens)
¶
Distinct-looking prompt token ids that stay within the vocabulary, which can be smaller than a full-budget or full-length prompt.
Source code in vllm/v1/worker/gpu/warmup.py
run_context_prefill_warmup(model_runner, worker_execute_model, worker_sample_tokens, num_context_tokens)
¶
Run a prefill step over about num_context_tokens of cached context.
MLA up-projects a prefix cache hit's context chunk by chunk, a transient that no step starting from an empty prefix reaches. The queries are kept short: the full-budget step already measured theirs.
Source code in vllm/v1/worker/gpu/warmup.py
run_kv_cache_sizing_warmup(model_runner, worker_execute_model, worker_sample_tokens)
¶
Run the steps an extensible KV cache measures its transient reserve on. These size rather than compile, so they run without JIT warmup.
Source code in vllm/v1/worker/gpu/warmup.py
run_mixed_prefill_decode_warmup(model_runner, worker_execute_model, worker_sample_tokens, num_tokens, *, mixed_step_context=None, req_id_prefix='_v2_mixed_warmup', prompt_logprobs=None, prefill_only_fallback=False)
¶
Run a V2 mixed prefill+decode step through normal scheduler inputs, requesting prompt_logprobs on the prefill requests.
Prefill tokens beyond what one request can hold (max_model_len - 1) are split across further prefill requests, up to max_num_reqs. If prefill_only_fallback, the prefills run alone when request slots rather than tokens limit them, including on a runner limited to one request.
Source code in vllm/v1/worker/gpu/warmup.py
164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 | |
warmup_kernels(model_runner, worker_execute_model, worker_sample_tokens)
¶
Run scheduler-realistic prefill and decode steps to JIT compile kernels.
We must call the provided worker's execute_model for pipeline parallel coordination.