anijain2305

TORCHDYNAMO_INLINE_INBUILT_NN_MODULES=0 TORCH_LOGS=recompiles python benchmarks/dynamo/torchbench.py --accuracy --no-translation-validation --inference --bfloat16 --backend eager --disable-cudagraphs --device cuda --only=hf_T5_generate
loading model: 0it [00:29, ?it/s]
cuda eval  hf_T5_generate
V0617 16:32:34.360000 140419876123776 torch/_dynamo/guards.py:2619] [8/1] [__recompiles] Recompiling function forward in /home/anijain/local/miniconda3/envs/pytorch2/lib/python3.11/site-packages/transformers/models/t5/modeling_t5.py:1639
V0617 16:32:34.360000 140419876123776 torch/_dynamo/guards.py:2619] [8/1] [__recompiles]     triggered by the following guard failure(s):
V0617 16:32:34.360000 140419876123776 torch/_dynamo/guards.py:2619] [8/1] [__recompiles]     - ___check_obj_id(L['past_key_values'], 8825760)
V0617 16:32:38.606000 140419876123776 torch/_dynamo/guards.py:2619] [10/1] [__recompiles] Recompiling function _update_model_kwargs_for_generation in /home/anijain/local/miniconda3/envs/pytorch2/lib/python3.11/

anijain2305

@torch.compile(backend="eager")
def fn(x, y, d):
    return x * y * d["foo"] * d["bar"]



V0410 15:48:57.778000 140318524949632 torch/_dynamo/guards.py:1785] [0/0] [__guards] GUARDS:
V0410 15:48:57.778000 140318524949632 torch/_dynamo/guards.py:1803] [0/0] [__guards] ___check_type_id(L['d'], 8833952)                             # return x * y * d["foo"] * d["bar"]  # examples/ord_dicts.py:24 in fn
V0410 15:48:57.778000 140318524949632 torch/_dynamo/guards.py:1803] [0/0] [__guards] len(L['d']) == 2                                              # return x * y * d["foo"] * d["bar"]  # examples/ord_dicts.py:24 in fn

anijain2305

import torch
from collections import OrderedDict

torch._dynamo.config.error_on_recompile = True


d = {
    30: 4,
    25: 2,
    20: 1,

anijain2305

import torch
torch._dynamo.config.guard_nn_modules = True

class SubMod(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = torch.nn.Linear(10, 10)

    def forward(self, x):
        return self.linear(x)

anijain2305

# minified2.py

import torch
from torch import nn
import os
from torch import distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
import torch._dynamo

# torch._dynamo.config.optimize_ddp = False

anijain2305

import torch
from torch.utils.checkpoint import checkpoint


class MyModule(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = torch.nn.Linear(10, 10)

    def _forward_helper(self, x):

anijain2305

import torch
import torchrec
from torchrec.sparse.jagged_tensor import _maybe_compute_kjt_to_jt_dict
from typing import List, Optional

torch._logging.set_logs(**torch._logging.DEFAULT_LOGGING)

anijain2305

******* loading model args.model='t5'
--> World Size = 1

--> Device_count = 2
--> running with these defaults train_config(seed=2023, verbose=True, total_steps_to_run=8, warmup_steps=5, use_orig_params=True, limit_all_gathers=True, use_ddp=False, ddp_bucket_size=25, ddp_use_gradient_view=False, hf_t5_checkpointing=False, print_memory_summary=False, print_training_loss_data=False, num_epochs=4, model_weights_bf16=False, use_mixed_precision=True, use_low_precision_gradient_policy=False, use_tf32=True, optimizer='AdamW', ap_use_kahan_summation=False, sharding_strategy=<ShardingStrategy.FULL_SHARD: 1>, print_sharding_plan=False, run_profiler=False, profile_folder='fsdp/profile_tracing', log_every=1, num_workers_dataloader=2, batch_size_training=16, fsdp_activation_checkpointing=True, use_fused_attention=False, use_parallel_attention=False, run_validation=True, memory_report=True, nccl_debug_handler=True, distributed_debug=True, use_non_recursive_wrapping=False, use_synthetic_data=False, use_deferred_init=False, 

anijain2305

import torch

class Foo(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.register_buffer("iterations", torch.tensor(0))

    def forward(self, x):
        self.iterations.add_(1)
        return x * self.iterations

anijain2305

import torch


class MockModule(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.a = torch.tensor(0)
        self.b = torch.tensor(0)
        self.register_buffer("iterations", torch.tensor(0))