TensorRT-LLM icon indicating copy to clipboard operation
TensorRT-LLM copied to clipboard
verified publisher icon NVIDIA

"Trying to remove block n by 0 that is not in hash map" spam in release 0.17

Open aikitoria opened this issue 10 months ago • 6 comments

System Info

  • 8x 4090 on dual Epyc server running Debian testing
  • CUDA toolkit version 12.8, driver version 570.86
  • Release container compiled from release 0.17 tag

Who can help?

Maybe @kaiyux or @byshiue

Information

  • [x] The official example scripts
  • [ ] My own modified scripts

Tasks

  • [x] An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
  • [ ] My own task or dataset (give details below)

Reproduction

  1. Convert a checkpoint
    python examples/llama/convert_checkpoint.py --model_dir /raid/models/raw/Mistral-Small-24B-Instruct-2501/ --tp_size 4 --dtype bfloat16 --output_dir /raid/models/checkpoint/Mistral-Small-24B-Instruct-2501_sm89_bf16_tp4 --workers 4

  2. Build an engine trtllm-build --checkpoint_dir /raid/models/checkpoint/Mistral-Small-24B-Instruct-2501_sm89_bf16_tp4/ --output_dir /raid/models/engine/Mistral-Small-24B-Instruct-2501_sm89_bf16_tp4 --max_batch_size 4 --workers 4 --gemm_plugin bfloat16 --use_paged_context_fmha enable

  3. Run the engine NCCL_P2P_LEVEL=SYS mpirun -n 4 --allow-run-as-root python examples/run.py --max_output_len 256 --engine_dir /raid/models/engine/Mistral-Small-24B-Instruct-2501_sm89_bf16_tp4/ --tokenizer_dir /raid/models/raw/Mistral-Small-24B-Instruct-2501/

Expected behavior

The engine runs normally

actual behavior

The engine runs, but there is log spam:

[TensorRT-LLM][INFO] [MemUsageChange] Allocated 9.27 GiB for max tokens in paged KV cache (243072).
[01/31/2025-05:22:09] [TRT-LLM] [I] Load engine takes: 4.808323860168457 sec
[01/31/2025-05:22:09] [TRT-LLM] [I] Load engine takes: 4.808316230773926 sec
[01/31/2025-05:22:09] [TRT-LLM] [I] Load engine takes: 4.808536767959595 sec
[01/31/2025-05:22:09] [TRT-LLM] [I] Load engine takes: 4.808875560760498 sec
[TensorRT-LLM][WARNING] Trying to remove block 1 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 2 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 3 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 4 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 2 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 3 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 4 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 2 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 3 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 4 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 2 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 3 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 4 by 0 that is not in hash map
Input [Text 0]: "<s>Born in north-east France, Soyer trained as a"

additional notes

This is a regression, it did not happen in 0.16 or the previous main snapshot.

When trying to use lookahead_decoding, the log gets spammed so much that it causes more overhead than was gained from the speculative decoding.

aikitoria avatar Jan 31 '25 05:01 aikitoria

This also occurs with trtllm-serve, not just run.py, which is much worse!

aikitoria avatar Jan 31 '25 05:01 aikitoria

Model config

{
    "version": "0.17.0.post1",
    "pretrained_config": {
        "mlp_bias": false,
        "attn_bias": false,
        "rotary_base": 500000.0,
        "rotary_scaling": {
            "factor": 16.0,
            "high_freq_factor": 4.0,
            "low_freq_factor": 1.0,
            "original_max_position_embeddings": 8192,
            "rope_type": "llama3"
        },
        "residual_mlp": false,
        "disable_weight_only_quant_plugin": false,
        "moe": {
            "num_experts": 0,
            "shared_expert_intermediate_size": 0,
            "top_k": 0,
            "normalization_mode": null,
            "sparse_mixer_epsilon": 0.01,
            "tp_mode": 0,
            "device_limited_n_group": 0,
            "device_limited_topk_group": 0,
            "device_limited_routed_scaling_factor": 1.0
        },
        "remove_duplicated_kv_heads": false,
        "fc_after_embed": false,
        "use_input_layernorm_in_first_layer": true,
        "use_last_layernorm": true,
        "layer_idx_offset": 0,
        "embedding_multiplier": 1.0,
        "attention_multiplier": 1.0,
        "residual_multiplier": 1.0,
        "output_multiplier_scale": 1.0,
        "architecture": "LlamaForCausalLM",
        "dtype": "bfloat16",
        "vocab_size": 128256,
        "hidden_size": 8192,
        "num_hidden_layers": 92,
        "num_attention_heads": 64,
        "hidden_act": "silu",
        "logits_dtype": "float32",
        "norm_epsilon": 1e-05,
        "runtime_defaults": null,
        "position_embedding_type": "rope_gpt_neox",
        "num_key_value_heads": 8,
        "intermediate_size": 32768,
        "max_position_embeddings": 131072,
        "mapping": {
            "world_size": 4,
            "gpus_per_node": 8,
            "cp_size": 1,
            "tp_size": 4,
            "pp_size": 1,
            "moe_tp_size": 4,
            "moe_ep_size": 1,
            "auto_parallel": false
        },
        "quantization": {
            "quant_algo": null,
            "kv_cache_quant_algo": null,
            "group_size": 128,
            "smoothquant_val": 0.5,
            "clamp_val": null,
            "use_meta_recipe": false,
            "has_zero_point": false,
            "pre_quant_scale": false,
            "exclude_modules": null
        },
        "use_parallel_embedding": false,
        "embedding_sharding_dim": 0,
        "head_size": 128,
        "qk_layernorm": false,
        "rotary_embedding_dim": 128,
        "tie_word_embeddings": false
    },
    "build_config": {
        "max_input_len": 130048,
        "max_seq_len": 131072,
        "opt_batch_size": 8,
        "max_batch_size": 256,
        "max_beam_width": 1,
        "max_num_tokens": 4096,
        "opt_num_tokens": 256,
        "max_prompt_embedding_table_size": 0,
        "kv_cache_type": "PAGED",
        "gather_context_logits": false,
        "gather_generation_logits": false,
        "strongly_typed": true,
        "force_num_profiles": null,
        "profiling_verbosity": "layer_names_only",
        "enable_debug_output": false,
        "max_draft_len": 0,
        "speculative_decoding_mode": 1,
        "use_refit": false,
        "input_timing_cache": null,
        "output_timing_cache": "model.cache",
        "lora_config": {
            "lora_dir": [],
            "lora_ckpt_source": "hf",
            "max_lora_rank": 64,
            "lora_target_modules": [],
            "trtllm_modules_to_hf_modules": {}
        },
        "auto_parallel_config": {
            "world_size": 1,
            "gpus_per_node": 8,
            "cluster_key": "H100-PCIe",
            "cluster_info": null,
            "sharding_cost_model": "alpha_beta",
            "comm_cost_model": "alpha_beta",
            "enable_pipeline_parallelism": false,
            "enable_shard_unbalanced_shape": false,
            "enable_shard_dynamic_shape": false,
            "enable_reduce_scatter": true,
            "builder_flags": null,
            "debug_mode": false,
            "infer_shape": true,
            "validation_mode": false,
            "same_buffer_io": {
                "past_key_value_(\\d+)": "present_key_value_\\1"
            },
            "same_spec_io": {},
            "sharded_io_allowlist": [
                "past_key_value_\\d+",
                "present_key_value_\\d*"
            ],
            "fill_weights": false,
            "parallel_config_cache": null,
            "profile_cache": null,
            "dump_path": null,
            "debug_outputs": []
        },
        "weight_sparsity": false,
        "weight_streaming": false,
        "plugin_config": {
            "dtype": "bfloat16",
            "bert_attention_plugin": "auto",
            "gpt_attention_plugin": "auto",
            "gemm_plugin": "bfloat16",
            "explicitly_disable_gemm_plugin": false,
            "gemm_swiglu_plugin": null,
            "fp8_rowwise_gemm_plugin": null,
            "qserve_gemm_plugin": null,
            "identity_plugin": null,
            "nccl_plugin": "bfloat16",
            "lora_plugin": null,
            "weight_only_groupwise_quant_matmul_plugin": null,
            "weight_only_quant_matmul_plugin": null,
            "smooth_quant_plugins": true,
            "smooth_quant_gemm_plugin": null,
            "layernorm_quantization_plugin": null,
            "rmsnorm_quantization_plugin": null,
            "quantize_per_token_plugin": false,
            "quantize_tensor_plugin": false,
            "moe_plugin": "auto",
            "mamba_conv1d_plugin": "auto",
            "low_latency_gemm_plugin": null,
            "low_latency_gemm_swiglu_plugin": null,
            "context_fmha": true,
            "bert_context_fmha_fp32_acc": false,
            "paged_kv_cache": true,
            "remove_input_padding": true,
            "reduce_fusion": false,
            "user_buffer": false,
            "tokens_per_block": 64,
            "use_paged_context_fmha": true,
            "use_fp8_context_fmha": false,
            "multiple_profiles": false,
            "paged_state": false,
            "streamingllm": false,
            "manage_weights": false,
            "use_fused_mlp": true,
            "pp_reduce_scatter": false
        },
        "use_strip_plan": false,
        "max_encoder_input_len": 1024,
        "monitor_memory": false,
        "use_mrope": false
    }
}

Run benchmark

CUDA_VISIBLE_DEVICES=4,5,6,7 mpirun -n 4 ./gptManagerBenchmark --engine_dir /data/models/4-gpu --dataset /datasets/cnn_dailymail.json --request_rate 5

Spam

....
[TensorRT-LLM][WARNING] Trying to remove block 1558 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1598 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1628 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1647 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1657 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1230 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1460 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1558 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1598 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1628 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1647 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1657 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1230 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1460 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1558 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1598 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1628 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1647 by 0 that is not in hash map
[TensorRT-LLM][WARNING] Trying to remove block 1657 by 0 that is not in hash map
....

akhoroshev avatar Feb 04 '25 10:02 akhoroshev

This warning can be ignored , will be fixed in next release.

chuangz0 avatar Feb 07 '25 06:02 chuangz0

You can assign me. Fix is on the way.

hypdeb avatar Feb 10 '25 15:02 hypdeb

Got same warnings when serving using triton backend

Harley-ZP avatar Feb 26 '25 08:02 Harley-ZP

This should be fixed now. Please verify and confirm if you can.

hypdeb avatar Mar 13 '25 14:03 hypdeb

I had this issue for my qwen2.5 14b model in 0.17 Does it means I need to upgrade to a newer version?

__

tlyhenry avatar Aug 14 '25 23:08 tlyhenry

Yes this has been fixed in releases > 0.17

hypdeb avatar Aug 15 '25 06:08 hypdeb