1 line
6.4 KiB
JSON
1 line
6.4 KiB
JSON
{"repo": "centaurus-alpha/unipool", "files": {"tests/functional_tests/test_cases/moe/gpt_dynamic_inference_tp4_etp1_pp1_ep8_16B_logitsmatch_zmq/model_config.yaml": {"9f57c44548323320750f298acddaebf024dcf9a2": {"t": 1785746844.1499395, "n": 3009, "h": "8392c90c686f2029", "c": "ENV_VARS:\n CUDA_DEVICE_MAX_CONNECTIONS: 1\n NVTE_ALLOW_NONDETERMINISTIC_ALGO: 0\n NCCL_ALGO: Ring\n CUBLAS_WORKSPACE_CONFIG: :4096:8\nTEST_TYPE: frozen-start\nMODE: inference\nMODEL_ARGS:\n --log-num-zeros-in-grad: true\n --log-validation-ppl-to-tensorboard: true\n --log-timers-to-tensorboard: true\n --log-memory-to-tensorboard: true\n --timing-log-level: 0\n --load: ${CHECKPOINT_LOAD_PATH}/model/deepseek_16b_pyt/dcp/mcore-v1_bf16/checkpoints\n --tokenizer-model: ${CHECKPOINT_LOAD_PATH}/model/deepseek_16b_pyt/dcp/mcore-v1_bf16/multiMixV8.gpt4o_nc_sd.500000.128k.vocab.json\n --tokenizer-type: TikTokenizer\n --tiktoken-pattern: v2\n --distributed-backend: nccl\n --log-interval: 1\n --transformer-impl: transformer_engine\n --tensor-model-parallel-size: 4\n --pipeline-model-parallel-size: 1\n --expert-model-parallel-size: 8\n --expert-tensor-parallel-size: 1\n --sequence-parallel: true\n --use-mcore-models: true\n --moe-token-dispatcher-type: alltoall\n --moe-grouped-gemm: true\n --num-experts: 64\n --moe-router-topk: 6\n --moe-z-loss-coeff: 0\n --moe-router-load-balancing-type: seq_aux_loss\n --moe-aux-loss-coeff: 1e-3\n --moe-router-score-function: sigmoid\n --untie-embeddings-and-output-weights: true\n --disable-bias-linear: true\n --init-method-std: 0.014\n --position-embedding-type: rope\n --rotary-base: 1000000\n --rotary-percent: 1.0\n --num-layers: 27\n --hidden-size: 2048\n --moe-ffn-hidden-size: 1408\n --moe-shared-expert-intermediate-size: 2816\n --ffn-hidden-size: 10944\n --num-attention-heads: 16\n --kv-channels: 128\n --normalization: RMSNorm\n --swiglu: true\n --attention-dropout: 0.0\n --hidden-dropout: 0.0\n --seq-length: 4096\n --max-position-embeddings: 4096\n --micro-batch-size: 1\n --ckpt-format: torch_dist\n --ckpt-fully-parallel-save: true\n --ckpt-fully-parallel-load: true\n --ckpt-assume-constant-structure: true\n --dist-ckpt-strictness: log_unexpected\n --bf16: true\n --attention-backend: flash\n --no-create-attention-mask-in-dataloader: true\n --num-workers: 8\n --use-checkpoint-args: true\n --no-use-tokenizer-model-from-checkpoint-args: true\n --no-load-optim: true\n --deterministic-mode: true # moe will use different ops for determinism for inference\n --save-interval: 2000\n --temperature: 1.0\n --top_k: 1\n --return-log-probs: true\n --num-tokens-to-generate: 30\n --max-tokens-to-oom: 3600000\n --inference-max-seq-length: 4096\n --output-path: ${TENSORBOARD_PATH}\n --prompts: \"Time travel to 2008, and go to a bar or a club or one of the myriad disco-basements on the Lower East Side that does not quite know which of those it is. Dance awkwardly in a room full of other glittered-up nerds, and wait for something to happen, buoyed on the feeling that this is the big swollen heart of life, that this is New York like the movies.\"\n --incoming-requests-per-sec: -1 # all requests arrive up front.\n --inference-repeat-n: 8\n --inference-logging-step-interval: 1 \n --sequence-parallel: true\n \nMETRICS:\n - \"generated_tokens\"\n - \"logprobs\"\n"}, "21dbd595bd371ee6757acff47c8e2a9e7af597cc": {"t": 1785768017.6649888, "n": 3009, "h": "8392c90c686f2029", "c": "ENV_VARS:\n CUDA_DEVICE_MAX_CONNECTIONS: 1\n NVTE_ALLOW_NONDETERMINISTIC_ALGO: 0\n NCCL_ALGO: Ring\n CUBLAS_WORKSPACE_CONFIG: :4096:8\nTEST_TYPE: frozen-start\nMODE: inference\nMODEL_ARGS:\n --log-num-zeros-in-grad: true\n --log-validation-ppl-to-tensorboard: true\n --log-timers-to-tensorboard: true\n --log-memory-to-tensorboard: true\n --timing-log-level: 0\n --load: ${CHECKPOINT_LOAD_PATH}/model/deepseek_16b_pyt/dcp/mcore-v1_bf16/checkpoints\n --tokenizer-model: ${CHECKPOINT_LOAD_PATH}/model/deepseek_16b_pyt/dcp/mcore-v1_bf16/multiMixV8.gpt4o_nc_sd.500000.128k.vocab.json\n --tokenizer-type: TikTokenizer\n --tiktoken-pattern: v2\n --distributed-backend: nccl\n --log-interval: 1\n --transformer-impl: transformer_engine\n --tensor-model-parallel-size: 4\n --pipeline-model-parallel-size: 1\n --expert-model-parallel-size: 8\n --expert-tensor-parallel-size: 1\n --sequence-parallel: true\n --use-mcore-models: true\n --moe-token-dispatcher-type: alltoall\n --moe-grouped-gemm: true\n --num-experts: 64\n --moe-router-topk: 6\n --moe-z-loss-coeff: 0\n --moe-router-load-balancing-type: seq_aux_loss\n --moe-aux-loss-coeff: 1e-3\n --moe-router-score-function: sigmoid\n --untie-embeddings-and-output-weights: true\n --disable-bias-linear: true\n --init-method-std: 0.014\n --position-embedding-type: rope\n --rotary-base: 1000000\n --rotary-percent: 1.0\n --num-layers: 27\n --hidden-size: 2048\n --moe-ffn-hidden-size: 1408\n --moe-shared-expert-intermediate-size: 2816\n --ffn-hidden-size: 10944\n --num-attention-heads: 16\n --kv-channels: 128\n --normalization: RMSNorm\n --swiglu: true\n --attention-dropout: 0.0\n --hidden-dropout: 0.0\n --seq-length: 4096\n --max-position-embeddings: 4096\n --micro-batch-size: 1\n --ckpt-format: torch_dist\n --ckpt-fully-parallel-save: true\n --ckpt-fully-parallel-load: true\n --ckpt-assume-constant-structure: true\n --dist-ckpt-strictness: log_unexpected\n --bf16: true\n --attention-backend: flash\n --no-create-attention-mask-in-dataloader: true\n --num-workers: 8\n --use-checkpoint-args: true\n --no-use-tokenizer-model-from-checkpoint-args: true\n --no-load-optim: true\n --deterministic-mode: true # moe will use different ops for determinism for inference\n --save-interval: 2000\n --temperature: 1.0\n --top_k: 1\n --return-log-probs: true\n --num-tokens-to-generate: 30\n --max-tokens-to-oom: 3600000\n --inference-max-seq-length: 4096\n --output-path: ${TENSORBOARD_PATH}\n --prompts: \"Time travel to 2008, and go to a bar or a club or one of the myriad disco-basements on the Lower East Side that does not quite know which of those it is. Dance awkwardly in a room full of other glittered-up nerds, and wait for something to happen, buoyed on the feeling that this is the big swollen heart of life, that this is New York like the movies.\"\n --incoming-requests-per-sec: -1 # all requests arrive up front.\n --inference-repeat-n: 8\n --inference-logging-step-interval: 1 \n --sequence-parallel: true\n \nMETRICS:\n - \"generated_tokens\"\n - \"logprobs\"\n"}}}} |