[BUG] 4-bit quantized models would repeatedly generate the same tokens when bf16.enabled is true about deepspeed

[Atry]

Workaround

If I change "bf16": {"enabled": True} to "fp16": {"enabled": True}, then the output would be:

Using quantizer for weights: CUDAQuantizer
[2024-06-10 21:48:13,828] [INFO] [partition_parameters.py:562:patch_init_and_builtins] Enable Zero3 engine with INT4 quantization.
[2024-06-10 21:48:14,225] [INFO] [partition_parameters.py:345:__exit__] finished initializing model - num_params = 603, num_elems = 3.30B
[2024-06-10 21:48:18,472] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Flops Profiler Enabled: False
[2024-06-10 21:48:18,473] [INFO] [logging.py:96:log_dist] [Rank 0] Creating ZeRO Offload
[2024-06-10 21:48:18,652] [INFO] [utils.py:779:see_memory_usage] DeepSpeedZeRoOffload initialize [begin]
[2024-06-10 21:48:18,653] [INFO] [utils.py:780:see_memory_usage] MA 1.78 GB         Max_MA 2.15 GB         CA 2.26 GB         Max_CA 2 GB 
[2024-06-10 21:48:18,654] [INFO] [utils.py:787:see_memory_usage] CPU Virtual Memory:  used = 7.4 GB, percent = 23.6%
Parameter Offload: Total persistent parameters: 92160 in 45 params
[2024-06-10 21:48:18,847] [INFO] [utils.py:779:see_memory_usage] DeepSpeedZeRoOffload initialize [end]
[2024-06-10 21:48:18,847] [INFO] [utils.py:780:see_memory_usage] MA 1.78 GB         Max_MA 1.78 GB         CA 2.26 GB         Max_CA 2 GB 
[2024-06-10 21:48:18,848] [INFO] [utils.py:787:see_memory_usage] CPU Virtual Memory:  used = 7.4 GB, percent = 23.6%
[2024-06-10 21:48:18,849] [INFO] [config.py:996:print] DeepSpeedEngine configuration:
[2024-06-10 21:48:18,850] [INFO] [config.py:1000:print]   activation_checkpointing_config  {
    "partition_activations": false, 
    "contiguous_memory_optimization": false, 
    "cpu_checkpointing": false, 
    "number_checkpoints": null, 
    "synchronize_checkpoint_boundary": false, 
    "profile": false
}
[2024-06-10 21:48:18,851] [INFO] [config.py:1000:print]   aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
[2024-06-10 21:48:18,852] [INFO] [config.py:1000:print]   amp_enabled .................. False
[2024-06-10 21:48:18,852] [INFO] [config.py:1000:print]   amp_params ................... False
[2024-06-10 21:48:18,853] [INFO] [config.py:1000:print]   autotuning_config ............ {
    "enabled": false, 
    "start_step": null, 
    "end_step": null, 
    "metric_path": null, 
    "arg_mappings": null, 
    "metric": "throughput", 
    "model_info": null, 
    "results_dir": "autotuning_results", 
    "exps_dir": "autotuning_exps", 
    "overwrite": true, 
    "fast": true, 
    "start_profile_step": 3, 
    "end_profile_step": 5, 
    "tuner_type": "gridsearch", 
    "tuner_early_stopping": 5, 
    "tuner_num_trials": 50, 
    "model_info_path": null, 
    "mp_size": 1, 
    "max_train_batch_size": null, 
    "min_train_batch_size": 1, 
    "max_train_micro_batch_size_per_gpu": 1.024000e+03, 
    "min_train_micro_batch_size_per_gpu": 1, 
    "num_tuning_micro_batch_sizes": 3
}
[2024-06-10 21:48:18,855] [INFO] [config.py:1000:print]   bfloat16_enabled ............. False
[2024-06-10 21:48:18,856] [INFO] [config.py:1000:print]   bfloat16_immediate_grad_update  False
[2024-06-10 21:48:18,857] [INFO] [config.py:1000:print]   checkpoint_parallel_write_pipeline  False
[2024-06-10 21:48:18,857] [INFO] [config.py:1000:print]   checkpoint_tag_validation_enabled  True
[2024-06-10 21:48:18,859] [INFO] [config.py:1000:print]   checkpoint_tag_validation_fail  False
[2024-06-10 21:48:18,859] [INFO] [config.py:1000:print]   comms_config ................. <deepspeed.comm.config.DeepSpeedCommsConfig object at 0x7f2b900e22d0>
[2024-06-10 21:48:18,860] [INFO] [config.py:1000:print]   communication_data_type ...... None
[2024-06-10 21:48:18,860] [INFO] [config.py:1000:print]   compile_config ............... enabled=False backend='inductor' kwargs={}
[2024-06-10 21:48:18,861] [INFO] [config.py:1000:print]   compression_config ........... {'weight_quantization': {'shared_parameters': {'enabled': False, 'quantizer_kernel': False, 'schedule_offset': 0, 'quantize_groups': 1, 'quantize_verbose': False, 'quantization_type': 'symmetric', 'quantize_weight_in_forward': False, 'rounding': 'nearest', 'fp16_mixed_quantize': False, 'quantize_change_ratio': 0.001}, 'different_groups': {}}, 'activation_quantization': {'shared_parameters': {'enabled': False, 'quantization_type': 'symmetric', 'range_calibration': 'dynamic', 'schedule_offset': 1000}, 'different_groups': {}}, 'sparse_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'row_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'head_pruning': {'shared_parameters': {'enabled': False, 'method': 'topk', 'schedule_offset': 1000}, 'different_groups': {}}, 'channel_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'layer_reduction': {'enabled': False}}
[2024-06-10 21:48:18,863] [INFO] [config.py:1000:print]   curriculum_enabled_legacy .... False
[2024-06-10 21:48:18,863] [INFO] [config.py:1000:print]   curriculum_params_legacy ..... False
[2024-06-10 21:48:18,864] [INFO] [config.py:1000:print]   data_efficiency_config ....... {'enabled': False, 'seed': 1234, 'data_sampling': {'enabled': False, 'num_epochs': 1000, 'num_workers': 0, 'curriculum_learning': {'enabled': False}}, 'data_routing': {'enabled': False, 'random_ltd': {'enabled': False, 'layer_token_lr_schedule': {'enabled': False}}}}
[2024-06-10 21:48:18,865] [INFO] [config.py:1000:print]   data_efficiency_enabled ...... False
[2024-06-10 21:48:18,865] [INFO] [config.py:1000:print]   dataloader_drop_last ......... False
[2024-06-10 21:48:18,866] [INFO] [config.py:1000:print]   disable_allgather ............ False
[2024-06-10 21:48:18,866] [INFO] [config.py:1000:print]   dump_state ................... False
[2024-06-10 21:48:18,867] [INFO] [config.py:1000:print]   dynamic_loss_scale_args ...... None
[2024-06-10 21:48:18,868] [INFO] [config.py:1000:print]   eigenvalue_enabled ........... False
[2024-06-10 21:48:18,868] [INFO] [config.py:1000:print]   eigenvalue_gas_boundary_resolution  1
[2024-06-10 21:48:18,869] [INFO] [config.py:1000:print]   eigenvalue_layer_name ........ bert.encoder.layer
[2024-06-10 21:48:18,869] [INFO] [config.py:1000:print]   eigenvalue_layer_num ......... 0
[2024-06-10 21:48:18,870] [INFO] [config.py:1000:print]   eigenvalue_max_iter .......... 100
[2024-06-10 21:48:18,871] [INFO] [config.py:1000:print]   eigenvalue_stability ......... 1e-06
[2024-06-10 21:48:18,871] [INFO] [config.py:1000:print]   eigenvalue_tol ............... 0.01
[2024-06-10 21:48:18,872] [INFO] [config.py:1000:print]   eigenvalue_verbose ........... False
[2024-06-10 21:48:18,872] [INFO] [config.py:1000:print]   elasticity_enabled ........... False
[2024-06-10 21:48:18,873] [INFO] [config.py:1000:print]   flops_profiler_config ........ {
    "enabled": false, 
    "recompute_fwd_factor": 0.0, 
    "profile_step": 1, 
    "module_depth": -1, 
    "top_modules": 1, 
    "detailed": true, 
    "output_file": null
}
[2024-06-10 21:48:18,874] [INFO] [config.py:1000:print]   fp16_auto_cast ............... False
[2024-06-10 21:48:18,874] [INFO] [config.py:1000:print]   fp16_enabled ................. True
[2024-06-10 21:48:18,875] [INFO] [config.py:1000:print]   fp16_master_weights_and_gradients  False
[2024-06-10 21:48:18,875] [INFO] [config.py:1000:print]   global_rank .................. 0
[2024-06-10 21:48:18,877] [INFO] [config.py:1000:print]   grad_accum_dtype ............. None
[2024-06-10 21:48:18,878] [INFO] [config.py:1000:print]   gradient_accumulation_steps .. 1
[2024-06-10 21:48:18,879] [INFO] [config.py:1000:print]   gradient_clipping ............ 0.0
[2024-06-10 21:48:18,879] [INFO] [config.py:1000:print]   gradient_predivide_factor .... 1.0
[2024-06-10 21:48:18,880] [INFO] [config.py:1000:print]   graph_harvesting ............. False
[2024-06-10 21:48:18,880] [INFO] [config.py:1000:print]   hybrid_engine ................ enabled=False max_out_tokens=512 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8
[2024-06-10 21:48:18,881] [INFO] [config.py:1000:print]   initial_dynamic_scale ........ 65536
[2024-06-10 21:48:18,882] [INFO] [config.py:1000:print]   load_universal_checkpoint .... False
[2024-06-10 21:48:18,883] [INFO] [config.py:1000:print]   loss_scale ................... 0
[2024-06-10 21:48:18,883] [INFO] [config.py:1000:print]   memory_breakdown ............. False
[2024-06-10 21:48:18,884] [INFO] [config.py:1000:print]   mics_hierarchial_params_gather  False
[2024-06-10 21:48:18,885] [INFO] [config.py:1000:print]   mics_shard_size .............. -1
[2024-06-10 21:48:18,885] [INFO] [config.py:1000:print]   monitor_config ............... tensorboard=TensorBoardConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') wandb=WandbConfig(enabled=False, group=None, team=None, project='deepspeed') csv_monitor=CSVConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') enabled=False
[2024-06-10 21:48:18,886] [INFO] [config.py:1000:print]   nebula_config ................ {
    "enabled": false, 
    "persistent_storage_path": null, 
    "persistent_time_interval": 100, 
    "num_of_version_in_retention": 2, 
    "enable_nebula_load": true, 
    "load_path": null
}
[2024-06-10 21:48:18,887] [INFO] [config.py:1000:print]   optimizer_legacy_fusion ...... False
[2024-06-10 21:48:18,887] [INFO] [config.py:1000:print]   optimizer_name ............... None
[2024-06-10 21:48:18,888] [INFO] [config.py:1000:print]   optimizer_params ............. None
[2024-06-10 21:48:18,889] [INFO] [config.py:1000:print]   pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0, 'pipe_partitioned': True, 'grad_partitioned': True}
[2024-06-10 21:48:18,890] [INFO] [config.py:1000:print]   pld_enabled .................. False
[2024-06-10 21:48:18,890] [INFO] [config.py:1000:print]   pld_params ................... False
[2024-06-10 21:48:18,891] [INFO] [config.py:1000:print]   prescale_gradients ........... False
[2024-06-10 21:48:18,895] [INFO] [config.py:1000:print]   scheduler_name ............... None
[2024-06-10 21:48:18,896] [INFO] [config.py:1000:print]   scheduler_params ............. None
[2024-06-10 21:48:18,896] [INFO] [config.py:1000:print]   seq_parallel_communication_data_type  torch.float32
[2024-06-10 21:48:18,897] [INFO] [config.py:1000:print]   sparse_attention ............. None
[2024-06-10 21:48:18,897] [INFO] [config.py:1000:print]   sparse_gradients_enabled ..... False
[2024-06-10 21:48:18,898] [INFO] [config.py:1000:print]   steps_per_print .............. 10
[2024-06-10 21:48:18,899] [INFO] [config.py:1000:print]   train_batch_size ............. 1
[2024-06-10 21:48:18,899] [INFO] [config.py:1000:print]   train_micro_batch_size_per_gpu  1
[2024-06-10 21:48:18,900] [INFO] [config.py:1000:print]   use_data_before_expert_parallel_  False
[2024-06-10 21:48:18,900] [INFO] [config.py:1000:print]   use_node_local_storage ....... False
[2024-06-10 21:48:18,901] [INFO] [config.py:1000:print]   wall_clock_breakdown ......... False
[2024-06-10 21:48:18,902] [INFO] [config.py:1000:print]   weight_quantization_config ... q_type='symmetric' q_groups=1 enabled=True num_bits=8 quantized_initialization={'num_bits': 4, 'group_size': 64, 'group_dim': 1, 'symmetric': False} post_init_quant={}
[2024-06-10 21:48:18,902] [INFO] [config.py:1000:print]   world_size ................... 1
[2024-06-10 21:48:18,903] [INFO] [config.py:1000:print]   zero_allow_untested_optimizer  False
[2024-06-10 21:48:18,904] [INFO] [config.py:1000:print]   zero_config .................. stage=3 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500,000,000 use_multi_rank_bucket_allreduce=True allgather_partitions=True allgather_bucket_size=500,000,000 overlap_comm=True load_from_fp32_weights=False elastic_checkpoint=False offload_param=None offload_optimizer=None sub_group_size=1,000,000,000 cpu_offload_param=None cpu_offload_use_pin_memory=None cpu_offload=None prefetch_bucket_size=50,000,000 param_persistence_threshold=100,000 model_persistence_threshold=sys.maxsize max_live_parameters=1,000,000,000 max_reuse_distance=1,000,000,000 gather_16bit_weights_on_model_save=False stage3_gather_fp16_weights_on_model_save=False ignore_unused_parameters=True legacy_stage1=False round_robin_gradients=False zero_hpz_partition_size=1 zero_quantized_weights=True zero_quantized_nontrainable_weights=True zero_quantized_gradients=False mics_shard_size=-1 mics_hierarchical_params_gather=False memory_efficient_linear=True pipeline_loading_checkpoint=False override_module_apply=True
[2024-06-10 21:48:18,904] [INFO] [config.py:1000:print]   zero_enabled ................. True
[2024-06-10 21:48:18,905] [INFO] [config.py:1000:print]   zero_force_ds_cpu_optimizer .. True
[2024-06-10 21:48:18,905] [INFO] [config.py:1000:print]   zero_optimization_stage ...... 3
[2024-06-10 21:48:18,906] [INFO] [config.py:986:print_user_config]   json = {
    "zero_optimization": {
        "load_from_fp32_weights": false, 
        "stage": 3, 
        "zero_quantized_weights": true, 
        "zero_quantized_nontrainable_weights": true
    }, 
    "train_micro_batch_size_per_gpu": 1, 
    "fp16": {
        "enabled": true
    }, 
    "weight_quantization": {
        "quantized_initialization": {
            "num_bits": 4, 
            "group_size": 64, 
            "group_dim": 1, 
            "symmetric": false
        }
    }
}
['<s>metro.org/metro-pilots-plan-to-build-new-metro-']

The generated text is now "metro.org/metro-pilots-plan-to-build-new-metro-". It does not repeat the same token any more.