| { |
| "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-4.7-Flash", |
| "base_revision": "7dd20894a642a0aa287e9827cb1a1f7f91386b67", |
| "fixture": "glm4-moe-lite-native-depth", |
| "variant": "fp8", |
| "parameter_counts": { |
| "architecture": "Glm4MoeLiteForCausalLM", |
| "backbone_parameters": 808008192, |
| "mtp_parameters": 13330944, |
| "total_parameters": 821339136, |
| "native_depth": true, |
| "model_patches": false, |
| "changes": { |
| "hidden_size": { |
| "original": 2048, |
| "tiny": 768 |
| }, |
| "intermediate_size": { |
| "original": 10240, |
| "tiny": 3072 |
| }, |
| "moe_intermediate_size": { |
| "original": 1536, |
| "tiny": 384 |
| }, |
| "num_attention_heads": { |
| "original": 20, |
| "tiny": 8 |
| }, |
| "num_key_value_heads": { |
| "original": 20, |
| "tiny": 8 |
| }, |
| "n_routed_experts": { |
| "original": 64, |
| "tiny": 8 |
| }, |
| "kv_lora_rank": { |
| "original": 512, |
| "tiny": 256 |
| }, |
| "q_lora_rank": { |
| "original": 768, |
| "tiny": 512 |
| }, |
| "n_group": { |
| "original": 1, |
| "tiny": 4 |
| }, |
| "topk_group": { |
| "original": 1, |
| "tiny": 4 |
| }, |
| "num_mtp_layers": { |
| "original": null, |
| "tiny": 1 |
| }, |
| "mtp_mlp_layer_types": { |
| "original": null, |
| "tiny": [ |
| "sparse" |
| ] |
| } |
| } |
| }, |
| "backbone_validation": { |
| "name": "glm4-moe-lite-native-depth", |
| "passed": true, |
| "reloaded_perplexity": 1.5342336153729887, |
| "model_patches": false, |
| "backbone_parameters": 808008192, |
| "backbone_activated_parameters_estimate": 645216768, |
| "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-lite-native-depth", |
| "world_size": 2, |
| "mtp_fp8_layers": 32, |
| "source": "fixtures/glm4-moe-lite-native-depth/source", |
| "quantized": "fixtures/glm4-moe-lite-native-depth/fp8" |
| }, |
| "vllm-source-results": { |
| "name": "glm4-moe-lite-native-depth", |
| "mode": "source", |
| "llm_args": { |
| "model": "fixtures/glm4-moe-lite-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": -1.0733059216290712, |
| "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": -0.7155595435760915, |
| "toy_keyword_match": true |
| }, |
| { |
| "prompt": "The classic game is called rock, paper,", |
| "text": " scissors. Rock beats scissors beats scissors beats paper, and paper beats rock. Rock", |
| "token_ids": [ |
| 71079, |
| 13, |
| 9124, |
| 33167, |
| 71079, |
| 33167, |
| 71079, |
| 33167, |
| 5567, |
| 11, |
| 323, |
| 5567, |
| 33167, |
| 6946, |
| 13, |
| 9124 |
| ], |
| "cumulative_logprob": -1.7960216077044606, |
| "toy_keyword_match": true |
| }, |
| { |
| "prompt": "Actions speak louder than", |
| "text": " words. What you do matters more what you say. say you do than what", |
| "token_ids": [ |
| 4244, |
| 13, |
| 3555, |
| 498, |
| 653, |
| 12844, |
| 803, |
| 1128, |
| 498, |
| 1977, |
| 13, |
| 1977, |
| 498, |
| 653, |
| 1091, |
| 1128 |
| ], |
| "cumulative_logprob": -12.015646234154701, |
| "toy_keyword_match": true |
| } |
| ], |
| "speculative_metrics": [], |
| "passed": true |
| }, |
| "vllm-fp8-results": { |
| "name": "glm4-moe-lite-native-depth", |
| "mode": "fp8", |
| "llm_args": { |
| "model": "fixtures/glm4-moe-lite-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": -1.0757355522364378, |
| "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": -0.7127121277153492, |
| "toy_keyword_match": true |
| }, |
| { |
| "prompt": "The classic game is called rock, paper,", |
| "text": " scissors. Rock beats scissors beats scissors beats paper, and paper beats rock. Rock", |
| "token_ids": [ |
| 71079, |
| 13, |
| 9124, |
| 33167, |
| 71079, |
| 33167, |
| 71079, |
| 33167, |
| 5567, |
| 11, |
| 323, |
| 5567, |
| 33167, |
| 6946, |
| 13, |
| 9124 |
| ], |
| "cumulative_logprob": -1.8343840935267508, |
| "toy_keyword_match": true |
| }, |
| { |
| "prompt": "Actions speak louder than", |
| "text": " words. What you do matters more what you say. say you do than what", |
| "token_ids": [ |
| 4244, |
| 13, |
| 3555, |
| 498, |
| 653, |
| 12844, |
| 803, |
| 1128, |
| 498, |
| 1977, |
| 13, |
| 1977, |
| 498, |
| 653, |
| 1091, |
| 1128 |
| ], |
| "cumulative_logprob": -11.956631049513817, |
| "toy_keyword_match": true |
| } |
| ], |
| "speculative_metrics": [], |
| "passed": true |
| }, |
| "vllm-mtp-results": { |
| "name": "glm4-moe-lite-native-depth", |
| "mode": "mtp", |
| "llm_args": { |
| "model": "fixtures/glm4-moe-lite-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": -1.0757355522364378, |
| "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": -0.7127121277153492, |
| "toy_keyword_match": true |
| }, |
| { |
| "prompt": "The classic game is called rock, paper,", |
| "text": " scissors. Rock beats scissors beats scissors beats paper, and paper beats rock. Rock", |
| "token_ids": [ |
| 71079, |
| 13, |
| 9124, |
| 33167, |
| 71079, |
| 33167, |
| 71079, |
| 33167, |
| 5567, |
| 11, |
| 323, |
| 5567, |
| 33167, |
| 6946, |
| 13, |
| 9124 |
| ], |
| "cumulative_logprob": -1.8343840935267508, |
| "toy_keyword_match": true |
| }, |
| { |
| "prompt": "Actions speak louder than", |
| "text": " words. What you do matters more what you say. say you do than what", |
| "token_ids": [ |
| 4244, |
| 13, |
| 3555, |
| 498, |
| 653, |
| 12844, |
| 803, |
| 1128, |
| 498, |
| 1977, |
| 13, |
| 1977, |
| 498, |
| 653, |
| 1091, |
| 1128 |
| ], |
| "cumulative_logprob": -11.956631049513817, |
| "toy_keyword_match": true |
| } |
| ], |
| "speculative_metrics": [ |
| { |
| "name": "vllm:spec_decode_num_drafts", |
| "value": 58 |
| }, |
| { |
| "name": "vllm:spec_decode_num_draft_tokens", |
| "value": 58 |
| }, |
| { |
| "name": "vllm:spec_decode_num_accepted_tokens", |
| "value": 3 |
| }, |
| { |
| "name": "vllm:spec_decode_num_accepted_tokens_per_pos", |
| "value": null |
| } |
| ], |
| "passed": true |
| } |
| } |
|
|
