assert benchmark times (#12042)

* assert jitted times in openpilot

* better error

* better error

* add ASSERT_MIN_STEP_TIME to more models

* t is step_times

* update benchmark times

* update times

examples/gpt2.py

@@ -181,6 +181,7 @@ class GPT2:
     self.tokenizer = tokenizer
 
   def generate(self, prompt:str, max_length:int, temperature:float, timing:bool=False, batch_size:int=1):
+    step_times = []
     prompt_tokens = self.tokenizer.encode(prompt, allowed_special={"<|endoftext|>"})
     toks = [prompt_tokens[:] for _ in range(batch_size)]
     start_pos = 0
@@ -197,8 +198,13 @@ class GPT2:
           else:
             tokens = Tensor([x[start_pos:] for x in toks])
           tok = self.model(tokens, Variable("start_pos", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()
+      step_times.append((GlobalCounters.time_sum_s-st)*1e3)
       start_pos = len(toks[0])
       for i,t in enumerate(tok): toks[i].append(t)
+
+    if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+      min_time = min(step_times)
+      assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
     return [self.tokenizer.decode(x) for x in toks]
 
 # **** main code ****

examples/hlb_cifar10.py

@@ -355,7 +355,7 @@ def train_cifar():
 
   # https://www.anandtech.com/show/16727/nvidia-announces-geforce-rtx-3080-ti-3070-ti-upgraded-cards-coming-in-june
   # 136 TFLOPS is the theoretical max w float16 on 3080 Ti
-
+  step_times = []
   model_ema: Optional[modelEMA] = None
   projected_ema_decay_val = hyp['ema']['decay_base'] ** hyp['ema']['every_n_steps']
   i = 0
@@ -413,12 +413,17 @@ def train_cifar():
           model_ema.update(model, Tensor([projected_ema_decay_val*(i/STEPS)**hyp['ema']['decay_pow']]))
 
       cl = time.monotonic()
+      step_times.append((cl-st)*1000.0)
       device_str = loss.device if isinstance(loss.device, str) else f"{loss.device[0]} * {len(loss.device)}"
       #  53  221.74 ms run,    2.22 ms python,  219.52 ms CL,  803.39 loss, 0.000807 LR, 4.66 GB used,   3042.49 GFLOPS,    674.65 GOPS
       print(f"{i:3d} {(cl-st)*1000.0:7.2f} ms run, {(et-st)*1000.0:7.2f} ms python, {(cl-et)*1000.0:7.2f} ms {device_str}, {loss_cpu:7.2f} loss, {opt_non_bias.lr.numpy()[0]:.6f} LR, {GlobalCounters.mem_used/1e9:.2f} GB used, {GlobalCounters.global_ops*1e-9/(cl-st):9.2f} GFLOPS, {GlobalCounters.global_ops*1e-9:9.2f} GOPS")
       st = cl
       i += 1
 
+  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+    min_time = min(step_times)
+    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
+
   # verify eval acc
   if target := getenv("TARGET_EVAL_ACC_PCT", 0.0):
     if eval_acc_pct >= target:

examples/mlperf/model_train.py

@@ -252,6 +252,10 @@ def train_resnet():
         print(f"epoch global_ops: {steps_in_train_epoch * GlobalCounters.global_ops:_}, "
               f"epoch global_mem: {steps_in_train_epoch * GlobalCounters.global_mem:_}")
         # if we are doing beam search, run the first eval too
+        if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+          min_time = min(step_times)
+          assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
+
         if (TRAIN_BEAM or EVAL_BEAM) and e == start_epoch: break
         return
     if MLLOGGER and RUNMLPERF:
@@ -344,6 +348,8 @@ def train_resnet():
         print(f"saving ckpt to {fn}")
         safe_save(get_training_state(model, optimizer_group, scheduler_group), fn)
 
+
+
 def train_retinanet():
   from contextlib import redirect_stdout
   from examples.mlperf.dataloader import batch_load_retinanet

examples/sdxl.py

@@ -6,7 +6,7 @@
 from tinygrad import Tensor, TinyJit, dtypes, GlobalCounters
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import safe_load, load_state_dict
-from tinygrad.helpers import fetch, trange, colored, Timing
+from tinygrad.helpers import fetch, trange, colored, Timing, getenv
 from extra.models.clip import Embedder, FrozenClosedClipEmbedder, FrozenOpenClipEmbedder
 from extra.models.unet import UNetModel, Upsample, Downsample, timestep_embedding
 from extra.bench_log import BenchEvent, WallTimeEvent
@@ -14,7 +14,7 @@ from examples.stable_diffusion import ResnetBlock, Mid
 import numpy as np
 
 from typing import Dict, List, Callable, Optional, Any, Set, Tuple, Union, Type
-import argparse, tempfile
+import argparse, tempfile, time
 from abc import ABC, abstractmethod
 from pathlib import Path
 from PIL import Image
@@ -342,11 +342,13 @@ class DPMPP2MSampler:
     sigmas = self.discretization(num_steps).to(x.device)
     x *= Tensor.sqrt(1.0 + sigmas[0] ** 2.0)
     num_sigmas = len(sigmas)
+    step_times = []
 
     old_denoised = None
     for i in trange(num_sigmas - 1):
       with Timing("step in ", enabled=timing, on_exit=lambda _: f", using {GlobalCounters.mem_used/1e9:.2f} GB"):
         GlobalCounters.reset()
+        st = time.perf_counter_ns()
         with WallTimeEvent(BenchEvent.STEP):
           x, old_denoised = self.sampler_step(
             old_denoised=old_denoised,
@@ -358,8 +360,13 @@ class DPMPP2MSampler:
             c=c,
             uc=uc,
           )
+          step_times.append(t:=(time.perf_counter_ns() - st)*1e-6)
           x.realize(old_denoised)
 
+    if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+      min_time = min(step_times)
+      assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
+
     return x
 
 

examples/stable_diffusion.py

@@ -2,7 +2,7 @@
 # https://github.com/ekagra-ranjan/huggingface-blog/blob/main/stable_diffusion.md
 import tempfile
 from pathlib import Path
-import argparse
+import argparse, time
 from collections import namedtuple
 from typing import Dict, Any
 
@@ -266,17 +266,23 @@ if __name__ == "__main__":
   def run(model, *x): return model(*x).realize()
 
   # this is diffusion
+  step_times = []
   with Context(BEAM=getenv("LATEBEAM")):
     for index, timestep in (t:=tqdm(list(enumerate(timesteps))[::-1])):
       GlobalCounters.reset()
+      st = time.perf_counter_ns()
       t.set_description("%3d %3d" % (index, timestep))
       with Timing("step in ", enabled=args.timing, on_exit=lambda _: f", using {GlobalCounters.mem_used/1e9:.2f} GB"):
         with WallTimeEvent(BenchEvent.STEP):
           tid = Tensor([index])
           latent = run(model, unconditional_context, context, latent, Tensor([timestep]), alphas[tid], alphas_prev[tid], Tensor([args.guidance]))
           if args.timing: Device[Device.DEFAULT].synchronize()
+      step_times.append((time.perf_counter_ns() - st)*1e-6)
     del run
 
+  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+    min_time = min(step_times)
+    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
   # upsample latent space to image with autoencoder
   x = model.decode(latent)
   print(x.shape)