| { |
| "mic": "0.3", |
| "name": "GLM-5.3-Flash multimodal generation", |
| "usage": "AutoProcessor + Glm5NextForConditionalGeneration, image+text chat, batched, causal LM labels", |
| "architecture": "Glm5NextForConditionalGeneration", |
| "checkpoints": { |
| "zai-org/GLM-5.3-Flash": { |
| "fixture": "huggingface-mic/GLM-5.3-Flash-mic" |
| } |
| }, |
| "fixtures": { |
| "huggingface-mic/GLM-5.3-Flash-mic": { |
| "built_from": "zai-org/GLM-5.3-Flash", |
| "covers": [ |
| "zai-org/GLM-5.3-Flash" |
| ], |
| "why_separate": "The checkpoint is a ~hundreds-of-billions-of-parameters FP8 MoE VLM; the slim fixture shrinks text depth/experts and vision depth so CPU runners can exercise the public forward, processor and training path.", |
| "build": { |
| "seed": 17, |
| "model_class": "Glm5NextForConditionalGeneration", |
| "config_overrides": { |
| "fields": { |
| "text_config.n_routed_experts": 8, |
| "text_config.num_experts_per_tok": 2, |
| "text_config.moe_intermediate_size": 64, |
| "text_config.intermediate_size": 128, |
| "text_config.hidden_size": 128, |
| "text_config.num_attention_heads": 4, |
| "text_config.num_key_value_heads": 4, |
| "text_config.q_lora_rank": 64, |
| "text_config.kv_lora_rank": 64, |
| "text_config.qk_nope_head_dim": 32, |
| "text_config.v_head_dim": 32, |
| "text_config.index_head_dim": 32, |
| "text_config.index_n_heads": 4, |
| "text_config.index_topk": 64, |
| "text_config.linear_attn_config.num_heads": 4, |
| "text_config.linear_attn_config.head_dim": 32, |
| "text_config.num_nextn_predict_layers": 1, |
| "vision_config.depth": 2, |
| "vision_config.hidden_size": 64, |
| "vision_config.intermediate_size": 128, |
| "vision_config.num_heads": 4, |
| "vision_config.out_hidden_size": 128, |
| "vision_config.image_size": 56, |
| "vision_config.patch_size": 14, |
| "vision_config.projection_intermediate_size": 128 |
| }, |
| "reason": "Shrink text and vision stacks to a few layers, few experts and a tiny vocabulary so the fixture fits CPU memory while preserving the MoE, linear-attention, sparse-attention, MTP and multimodal projector code paths." |
| } |
| } |
| } |
| }, |
| "integrations": { |
| "transformers": { |
| "entry_point": "integrations/transformers/usage.md", |
| "format": "runnable-markdown", |
| "requires": [ |
| "transformers>=5", |
| "torch", |
| "torchvision" |
| ], |
| "targets": [ |
| "fixtures" |
| ], |
| "variants": { |
| "default": {}, |
| "eager": { |
| "attn_implementation": "eager" |
| }, |
| "bfloat16": { |
| "dtype": "bfloat16" |
| } |
| }, |
| "training": true, |
| "resources": { |
| "fixture": { |
| "memory_gb": 4, |
| "download_mb": 50 |
| }, |
| "zai-org/GLM-5.3-Flash": { |
| "devices": [ |
| "cuda" |
| ], |
| "gpus": 1, |
| "gpu_memory_gb": 80, |
| "memory_gb": 64, |
| "disk_gb": 400, |
| "download_mb": 300000, |
| "variants": { |
| "bfloat16": { |
| "gpu_memory_gb": 80 |
| }, |
| "default": { |
| "devices": [] |
| }, |
| "eager": { |
| "devices": [] |
| } |
| } |
| } |
| } |
| } |
| }, |
| "outputs": { |
| "runtime": { |
| "compare": "exact" |
| }, |
| "predicted_tokens": { |
| "compare": "exact" |
| }, |
| "answer_logits": { |
| "compare": "allclose", |
| "atol": 0.0001, |
| "rtol": 0.0001 |
| }, |
| "training.loss": { |
| "compare": "allclose", |
| "atol": 0.0001, |
| "rtol": 0.0001 |
| }, |
| "training.gradients": { |
| "compare": "all_true" |
| }, |
| "training.updated": { |
| "compare": "true" |
| } |
| } |
| } |
|
|
