File size: 13,483 Bytes
4fbc241 1d6826f 4fbc241 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 | from __future__ import annotations
import random
from typing import Any
from llmserve_env.models import MetricsSnapshot, QuantizationTier, ServeAction, WorkloadSnapshot
from llmserve_env.task_catalog import get_task_config
from server.kv_cache_simulator import KVCacheSimulator
from server.replay_assets import load_lookup_table
from server.speculative_decoder import SpeculativeDecoder
class TraceSimulator:
def __init__(self, seed: int = 42) -> None:
self.seed = seed
try:
self.kv_cache = KVCacheSimulator()
except Exception as e:
raise RuntimeError(f"Failed to initialize KVCacheSimulator: {e}") from e
try:
self.speculative_decoder = SpeculativeDecoder()
except Exception as e:
raise RuntimeError(f"Failed to initialize SpeculativeDecoder: {e}") from e
try:
self.lookup_table = load_lookup_table("lookup_tables/latency_table.parquet")
except Exception as e:
raise RuntimeError(f"Failed to load lookup table: {e}") from e
self.batch_history: list[int] = []
def reset(self, seed: int | None = None) -> None:
if seed is not None:
self.seed = seed
self.batch_history = []
def simulate_step(self, task_id: str, action: ServeAction, workload: WorkloadSnapshot) -> MetricsSnapshot:
task = get_task_config(task_id)
batch_effective = min(action.batch_cap, max(1, workload.queue_depth + int(workload.arrival_rate)))
quantization_tier = QuantizationTier(action.quantization_tier)
self.batch_history.append(action.batch_cap)
if len(self.batch_history) > 10:
self.batch_history.pop(0)
is_throttled = False
if workload.step_index > 100 and len(self.batch_history) == 10:
avg_batch = sum(self.batch_history) / 10.0
if avg_batch > 0.8 * 512:
is_throttled = True
profile = self._lookup_profile(
quantization_tier=quantization_tier,
batch_effective=batch_effective,
prompt_length=workload.mean_prompt_length,
kv_budget_fraction=action.kv_budget_fraction,
speculation_depth=action.speculation_depth,
)
kv_occupancy, evictions = self.kv_cache.apply(
queue_depth=workload.queue_depth,
mean_prompt_length=workload.mean_prompt_length,
kv_budget_fraction=action.kv_budget_fraction,
priority_routing=action.priority_routing,
)
spec_acceptance, itl_speedup = self.speculative_decoder.estimate(
task_id=task_id,
speculation_depth=action.speculation_depth,
mean_prompt_length=workload.mean_prompt_length,
)
queue_pressure = 1.0 + min(2.5, workload.queue_depth / max(1.0, float(action.batch_cap))) * 0.18
arrival_pressure = 1.0 + min(1.5, workload.arrival_rate / max(1.0, float(action.batch_cap))) * 0.04
phase_factor = {"warmup": 1.03, "steady": 1.0, "burst": 1.10, "cooldown": 0.98}.get(workload.phase, 1.0)
throughput_noise = self._noise(task_id, action, workload, "throughput")
latency_noise = self._noise(task_id, action, workload, "latency")
memory_noise = self._noise(task_id, action, workload, "memory")
cost_noise = self._noise(task_id, action, workload, "cost")
recomp_drop = evictions * max(1.0, workload.mean_prompt_length / 128.0) * 0.02
throughput_tps = max(
1.0,
(profile["throughput_tps"] - recomp_drop)
* throughput_noise
* profile["quantization_speedup"]
* (1.0 + (0.05 if action.prefill_decode_split else 0.0))
* (1.0 + (0.03 if action.priority_routing and workload.priority_fraction > 0 else 0.0))
* (1.0 + spec_acceptance * 0.18)
/ max(1.0, queue_pressure * 0.35),
)
thermal_mult = 1.15 if is_throttled else 1.0
p50_ttft_ms = max(
20.0,
profile["p50_ttft_ms"]
* latency_noise
* queue_pressure
* arrival_pressure
* phase_factor
* thermal_mult
* profile["quantization_latency_mult"]
* (0.92 if action.prefill_decode_split else 1.0)
* (1.0 - min(0.22, spec_acceptance * 0.25)),
)
p99_ttft_ms = max(
p50_ttft_ms,
profile["p99_ttft_ms"]
* latency_noise
* queue_pressure
* arrival_pressure
* phase_factor
* thermal_mult
* profile["quantization_latency_mult"]
* (1.0 - min(0.16, spec_acceptance * 0.18))
* (1.0 + kv_occupancy * 0.08),
)
if task_id == "adversarial_multitenant" and workload.phase == "mega-prompt" and not action.prefill_decode_split:
p99_ttft_ms *= 5.0
p50_itl_ms = max(
1.5,
profile["p50_itl_ms"]
* itl_speedup
* thermal_mult
* profile["quantization_itl_mult"]
* (1.0 + kv_occupancy * 0.08)
* self._noise(task_id, action, workload, "itl"),
)
gpu_memory_used_gb = max(
2.0,
(
profile["gpu_memory_gb"] * profile["quantization_memory_mult"]
+ kv_occupancy * 6.5
+ workload.mean_prompt_length / 2200.0
+ workload.queue_depth / 140.0
- (0.7 if action.priority_routing else 0.0)
)
* memory_noise,
)
estimated_cost_per_1k = max(
0.0003,
(
profile["base_cost_per_1k"] * profile["quantization_cost_mult"]
* (1.0 + kv_occupancy * 0.12)
* (1.0 + (0.04 if action.prefill_decode_split else 0.0))
)
* cost_noise,
)
requests_served = min(batch_effective, max(0, workload.queue_depth + int(workload.arrival_rate)))
slo_violations = 0
if gpu_memory_used_gb > float(task["memory_cap_gb"]):
gpu_memory_used_gb = float(task["memory_cap_gb"])
evictions += max(1, batch_effective // 5)
slo_violations += max(1, batch_effective // 6)
return MetricsSnapshot(
p50_ttft_ms=p50_ttft_ms,
p99_ttft_ms=p99_ttft_ms,
p50_itl_ms=p50_itl_ms,
throughput_tps=throughput_tps,
gpu_memory_used_gb=gpu_memory_used_gb,
estimated_cost_per_1k=estimated_cost_per_1k,
spec_acceptance_rate=spec_acceptance,
eviction_events=evictions,
preemption_events=int(evictions if action.priority_routing and kv_occupancy > 0.95 else 0),
is_throttled=is_throttled,
slo_violations=slo_violations,
requests_served=requests_served,
)
def _lookup_profile(
self,
quantization_tier: QuantizationTier,
batch_effective: int,
prompt_length: float,
kv_budget_fraction: float,
speculation_depth: int,
) -> dict[str, float]:
prompt_bucket = _prompt_size_bucket(prompt_length)
batch_points = sorted(int(value) for value in self.lookup_table["batch_cap_bucket"].unique())
kv_points = sorted(float(value) for value in self.lookup_table["kv_budget_bucket"].unique())
spec_points = sorted(int(value) for value in self.lookup_table["spec_depth_bucket"].unique())
batch_low, batch_high = _bounding_points(batch_points, batch_effective)
kv_low, kv_high = _bounding_points(kv_points, kv_budget_fraction)
spec_low, spec_high = _bounding_points(spec_points, speculation_depth)
corners: list[tuple[dict[str, Any], float]] = []
for batch_bucket, batch_weight in ((batch_low, 1.0 - _interpolation_weight(batch_low, batch_high, batch_effective)), (batch_high, _interpolation_weight(batch_low, batch_high, batch_effective))):
for kv_bucket, kv_weight in ((kv_low, 1.0 - _interpolation_weight(kv_low, kv_high, kv_budget_fraction)), (kv_high, _interpolation_weight(kv_low, kv_high, kv_budget_fraction))):
for spec_bucket, spec_weight in ((spec_low, 1.0 - _interpolation_weight(spec_low, spec_high, speculation_depth)), (spec_high, _interpolation_weight(spec_low, spec_high, speculation_depth))):
weight = batch_weight * kv_weight * spec_weight
if weight <= 0.0:
continue
row = self._nearest_row(
batch_bucket=batch_bucket,
kv_bucket=kv_bucket,
spec_bucket=spec_bucket,
prompt_bucket=prompt_bucket,
)
corners.append((row, weight))
if not corners:
corners.append((self._nearest_row(batch_effective, kv_budget_fraction, speculation_depth, prompt_bucket), 1.0))
metrics = ["throughput_tps", "p50_ttft_ms", "p99_ttft_ms", "p50_itl_ms", "gpu_memory_gb"]
profile = {metric: 0.0 for metric in metrics}
total_weight = sum(weight for _, weight in corners) or 1.0
for row, weight in corners:
normalized_weight = weight / total_weight
for metric in metrics:
profile[metric] += float(row[metric]) * normalized_weight
if speculation_depth > 0 and not any(int(row["spec_depth_bucket"]) == speculation_depth for row, _ in corners):
depth_factor = 1.0 + min(0.18, speculation_depth * 0.025)
profile["throughput_tps"] *= depth_factor
profile["p50_ttft_ms"] *= max(0.75, 1.0 - speculation_depth * 0.03)
profile["p99_ttft_ms"] *= max(0.78, 1.0 - speculation_depth * 0.025)
profile["p50_itl_ms"] *= max(0.78, 1.0 - speculation_depth * 0.02)
profile["gpu_memory_gb"] *= 1.0 + speculation_depth * 0.015
quantization_profiles = {
QuantizationTier.FP16: {
"quantization_speedup": 1.00,
"quantization_latency_mult": 1.00,
"quantization_itl_mult": 1.00,
"quantization_memory_mult": 1.00,
"quantization_cost_mult": 1.00,
},
QuantizationTier.INT8: {
"quantization_speedup": 1.08,
"quantization_latency_mult": 0.94,
"quantization_itl_mult": 0.94,
"quantization_memory_mult": 0.82,
"quantization_cost_mult": 0.78,
},
QuantizationTier.INT4: {
"quantization_speedup": 1.16,
"quantization_latency_mult": 0.90,
"quantization_itl_mult": 0.90,
"quantization_memory_mult": 0.68,
"quantization_cost_mult": 0.62,
},
}
profile.update(quantization_profiles[quantization_tier])
profile["base_cost_per_1k"] = max(0.0004, (profile["gpu_memory_gb"] * 0.0012 + batch_effective * 0.000003) / max(profile["throughput_tps"], 1.0) * 1000.0)
return profile
def _nearest_row(
self,
batch_bucket: int,
kv_bucket: float,
spec_bucket: int,
prompt_bucket: str,
) -> dict[str, Any]:
frame = self.lookup_table[self.lookup_table["prompt_size_bucket"] == prompt_bucket]
if frame.empty:
frame = self.lookup_table
distance_frame = frame.assign(
_distance=(
(frame["batch_cap_bucket"].astype(float) - float(batch_bucket)).abs() / 256.0
+ (frame["kv_budget_bucket"].astype(float) - float(kv_bucket)).abs() * 2.0
+ (frame["spec_depth_bucket"].astype(float) - float(spec_bucket)).abs() / 8.0
)
)
row = distance_frame.sort_values(["_distance", "batch_cap_bucket"]).iloc[0]
return row.to_dict()
def _noise(
self,
task_id: str,
action: ServeAction,
workload: WorkloadSnapshot,
metric: str,
) -> float:
seed_material = (
f"{self.seed}|{task_id}|{metric}|{workload.phase}|{workload.step_index}|"
f"{workload.queue_depth}|{round(workload.arrival_rate, 3)}|{round(workload.mean_prompt_length, 3)}"
)
rng = random.Random(seed_material)
sigma = 0.03
if action.quantization_tier in [QuantizationTier.INT8.value, QuantizationTier.INT4.value]:
sigma = 0.07
if workload.phase in ["burst", "mega-prompt"]:
sigma = 0.12
return float(rng.gauss(1.0, sigma))
def _bounding_points(points: list[float], value: float) -> tuple[float, float]:
lower = points[0]
upper = points[-1]
for point in points:
if point <= value:
lower = point
if point >= value:
upper = point
break
return lower, upper
def _interpolation_weight(lower: float, upper: float, value: float) -> float:
if upper == lower:
return 0.0
return max(0.0, min(1.0, (value - lower) / (upper - lower)))
def _lerp(start: float, end: float, weight: float) -> float:
return start + (end - start) * weight
def _prompt_size_bucket(prompt_length: float) -> str:
if prompt_length <= 256:
return "small"
if prompt_length <= 2048:
return "medium"
return "large"
|