{ "llm_compressor_commit": "2d5242056a8a00028686a303771ffc8e2fe27d07", "compressed_tensors_commit": "e69c8dc58aa152e5f5e36e85800ec1d2e7de5271", "transformers": "5.17.0", "torch": "2.14.0+cu130", "vllm": "0.30.0", "base_architecture": "zai-org/GLM-4.5-Air", "base_revision": "a24ceef6ce4f3536971efe9b778bdaa1bab18daa", "fixture": "glm4-moe-native-depth", "variant": "source", "parameter_counts": { "architecture": "Glm4MoeForCausalLM", "backbone_parameters": 707149568, "mtp_parameters": 11513600, "total_parameters": 718663168, "native_depth": true, "model_patches": false, "changes": { "hidden_size": { "original": 4096, "tiny": 768 }, "intermediate_size": { "original": 10944, "tiny": 3072 }, "num_attention_heads": { "original": 96, "tiny": 8 }, "num_key_value_heads": { "original": 8, "tiny": 4 }, "moe_intermediate_size": { "original": 1408, "tiny": 384 }, "num_experts_per_tok": { "original": 8, "tiny": 4 }, "n_routed_experts": { "original": 128, "tiny": 8 } } }, "backbone_validation": { "name": "glm4-moe-native-depth", "passed": true, "reloaded_perplexity": 1.1919350007360576, "model_patches": false, "backbone_parameters": 707149568, "backbone_activated_parameters_estimate": 547897088, "generation": "The capital of France is Paris. Paris is the largest city in France and serves as the country's political" }, "mtp_training": "synthetic initialized projections with trained final decoder block copied; not separately trained", "training": { "seed": 3225, "initialization": "random", "optimizer": "AdamW", "learning_rate": 0.0004, "weight_decay": 0.01, "batch_size": 2, "max_sequence_length": 160, "evaluation": "same toy corpus as training" }, "quantization-results": { "name": "glm4-moe-native-depth", "world_size": 2, "mtp_fp8_layers": 31, "source": "fixtures/glm4-moe-native-depth/source", "quantized": "fixtures/glm4-moe-native-depth/fp8" }, "vllm-source-results": { "name": "glm4-moe-native-depth", "mode": "source", "llm_args": { "model": "fixtures/glm4-moe-native-depth/source", "dtype": "bfloat16", "tensor_parallel_size": 1, "gpu_memory_utilization": 0.25, "max_model_len": 256, "max_num_seqs": 4, "enforce_eager": true, "enable_prefix_caching": false, "seed": 3225, "disable_log_stats": false }, "generations": [ { "prompt": "The capital of France is", "text": " Paris. Paris is the largest city in France and serves as the country's political", "token_ids": [ 12089, 13, 12089, 374, 279, 7772, 3283, 304, 9621, 323, 17026, 438, 279, 3146, 594, 4948 ], "cumulative_logprob": -0.30267810029909015, "toy_keyword_match": true }, { "prompt": "The creator of the theory of relativity was Albert", "text": " Einstein. Einstein developed the special theory of relativity in 1905 and the", "token_ids": [ 53690, 13, 53690, 7881, 279, 3281, 10121, 315, 1351, 43166, 304, 220, 114146, 20, 323, 279 ], "cumulative_logprob": -0.2019934607669711, "toy_keyword_match": true }, { "prompt": "The classic game is called rock, paper,", "text": " scissors. Rock beats scissors, scissors beats paper, and paper beats rock. rock", "token_ids": [ 71079, 13, 9124, 33167, 71079, 11, 71079, 33167, 5567, 11, 323, 5567, 33167, 6946, 13, 6946 ], "cumulative_logprob": -1.577314083231613, "toy_keyword_match": true }, { "prompt": "Actions speak louder than", "text": " What you do matters more than. you say. you say. you do matters", "token_ids": [ 3555, 498, 653, 12844, 803, 1091, 13, 498, 1977, 13, 498, 1977, 13, 498, 653, 12844 ], "cumulative_logprob": -3.4432916017249227, "toy_keyword_match": false } ], "speculative_metrics": [], "passed": true }, "vllm-fp8-results": { "name": "glm4-moe-native-depth", "mode": "fp8", "llm_args": { "model": "fixtures/glm4-moe-native-depth/fp8", "dtype": "bfloat16", "tensor_parallel_size": 1, "gpu_memory_utilization": 0.25, "max_model_len": 256, "max_num_seqs": 4, "enforce_eager": true, "enable_prefix_caching": false, "seed": 3225, "disable_log_stats": false }, "generations": [ { "prompt": "The capital of France is", "text": " Paris. Paris is the largest city in France and serves as the country's political", "token_ids": [ 12089, 13, 12089, 374, 279, 7772, 3283, 304, 9621, 323, 17026, 438, 279, 3146, 594, 4948 ], "cumulative_logprob": -0.30267687886953354, "toy_keyword_match": true }, { "prompt": "The creator of the theory of relativity was Albert", "text": " Einstein. Einstein developed the special theory of relativity in 1905 and the", "token_ids": [ 53690, 13, 53690, 7881, 279, 3281, 10121, 315, 1351, 43166, 304, 220, 114146, 20, 323, 279 ], "cumulative_logprob": -0.20183802442625165, "toy_keyword_match": true }, { "prompt": "The classic game is called rock, paper,", "text": " scissors. Rock beats scissors, scissors beats paper, and paper beats rock. rock", "token_ids": [ 71079, 13, 9124, 33167, 71079, 11, 71079, 33167, 5567, 11, 323, 5567, 33167, 6946, 13, 6946 ], "cumulative_logprob": -1.585465795127675, "toy_keyword_match": true }, { "prompt": "Actions speak louder than", "text": " What you do matters more than. you say. you say. you do matters", "token_ids": [ 3555, 498, 653, 12844, 803, 1091, 13, 498, 1977, 13, 498, 1977, 13, 498, 653, 12844 ], "cumulative_logprob": -3.5015990268439054, "toy_keyword_match": false } ], "speculative_metrics": [], "passed": true }, "vllm-mtp-results": { "name": "glm4-moe-native-depth", "mode": "mtp", "llm_args": { "model": "fixtures/glm4-moe-native-depth/fp8", "dtype": "bfloat16", "tensor_parallel_size": 1, "gpu_memory_utilization": 0.25, "max_model_len": 256, "max_num_seqs": 4, "enforce_eager": true, "enable_prefix_caching": false, "seed": 3225, "disable_log_stats": false, "speculative_config": { "method": "mtp", "num_speculative_tokens": 1 } }, "generations": [ { "prompt": "The capital of France is", "text": " Paris. Paris is the largest city in France and serves as the country's political", "token_ids": [ 12089, 13, 12089, 374, 279, 7772, 3283, 304, 9621, 323, 17026, 438, 279, 3146, 594, 4948 ], "cumulative_logprob": -0.30267687886953354, "toy_keyword_match": true }, { "prompt": "The creator of the theory of relativity was Albert", "text": " Einstein. Einstein developed the special theory of relativity in 1905 and the", "token_ids": [ 53690, 13, 53690, 7881, 279, 3281, 10121, 315, 1351, 43166, 304, 220, 114146, 20, 323, 279 ], "cumulative_logprob": -0.20183802442625165, "toy_keyword_match": true }, { "prompt": "The classic game is called rock, paper,", "text": " scissors. Rock beats scissors, scissors beats paper, and paper beats rock. rock", "token_ids": [ 71079, 13, 9124, 33167, 71079, 11, 71079, 33167, 5567, 11, 323, 5567, 33167, 6946, 13, 6946 ], "cumulative_logprob": -1.585465795127675, "toy_keyword_match": true }, { "prompt": "Actions speak louder than", "text": " What you do matters more than. you say. you say. you do matters", "token_ids": [ 3555, 498, 653, 12844, 803, 1091, 13, 498, 1977, 13, 498, 1977, 13, 498, 653, 12844 ], "cumulative_logprob": -3.5015990268439054, "toy_keyword_match": false } ], "speculative_metrics": [ { "name": "vllm:spec_decode_num_drafts", "value": 60 }, { "name": "vllm:spec_decode_num_draft_tokens", "value": 60 }, { "name": "vllm:spec_decode_num_accepted_tokens", "value": 0 }, { "name": "vllm:spec_decode_num_accepted_tokens_per_pos", "value": null } ], "passed": true } }