tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
1088c4f verified | { | |
| "mic": "0.3", | |
| "name": "GLM-MoE-DSA causal LM", | |
| "usage": "AutoTokenizer + AutoModelForCausalLM, batched, left padding, next-token loss", | |
| "architecture": "GlmMoeDsaForCausalLM", | |
| "checkpoints": { | |
| "zai-org/GLM-5.3": { | |
| "fixture": "huggingface-mic/GLM-5.3-mic" | |
| } | |
| }, | |
| "fixtures": { | |
| "huggingface-mic/GLM-5.3-mic": { | |
| "built_from": "zai-org/GLM-5.3", | |
| "covers": [ | |
| "zai-org/GLM-5.3" | |
| ], | |
| "why_separate": "The checkpoint is a 78-layer 256-expert MoE with FP8 block-quantized weights; the slim fixture shrinks depth, experts, and vocab so CPU runners can measure the public API without the full download.", | |
| "build": { | |
| "seed": 17, | |
| "model_class": "GlmMoeDsaForCausalLM", | |
| "config_overrides": { | |
| "fields": { | |
| "n_routed_experts": 8, | |
| "n_shared_experts": 1, | |
| "num_experts_per_tok": 2, | |
| "moe_intermediate_size": 64, | |
| "intermediate_size": 128, | |
| "hidden_size": 128, | |
| "num_attention_heads": 4, | |
| "num_key_value_heads": 4, | |
| "head_dim": 32, | |
| "q_lora_rank": 64, | |
| "kv_lora_rank": 32, | |
| "qk_nope_head_dim": 24, | |
| "qk_rope_head_dim": 8, | |
| "v_head_dim": 32, | |
| "index_head_dim": 16, | |
| "index_n_heads": 2, | |
| "index_topk": 8, | |
| "first_k_dense_replace": 1, | |
| "num_nextn_predict_layers": 1, | |
| "max_position_embeddings": 512, | |
| "dtype": "float32" | |
| }, | |
| "reason": "Shrink depth, expert count, and vocab so the fixture fits CPU memory; shrink already drops the FP8 quantization_config, so fixture weights are plain float32." | |
| } | |
| } | |
| } | |
| }, | |
| "integrations": { | |
| "transformers": { | |
| "entry_point": "integrations/transformers/usage.md", | |
| "format": "runnable-markdown", | |
| "requires": [ | |
| "transformers>=5", | |
| "torch" | |
| ], | |
| "targets": [ | |
| "fixtures" | |
| ], | |
| "variants": { | |
| "default": {}, | |
| "eager": { | |
| "attn_implementation": "eager" | |
| }, | |
| "bfloat16": { | |
| "dtype": "bfloat16" | |
| } | |
| }, | |
| "training": true, | |
| "resources": { | |
| "fixture": { | |
| "memory_gb": 2, | |
| "download_mb": 20 | |
| }, | |
| "zai-org/GLM-5.3": { | |
| "devices": [ | |
| "cuda" | |
| ], | |
| "gpus": 1, | |
| "gpu_memory_gb": 80, | |
| "memory_gb": 64, | |
| "download_mb": 350000 | |
| } | |
| } | |
| } | |
| }, | |
| "outputs": { | |
| "runtime": { | |
| "compare": "exact" | |
| }, | |
| "next_token_ids": { | |
| "compare": "exact" | |
| }, | |
| "next_token_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" | |
| } | |
| } | |
| } | |