| { | |
| "mic": "0.3", | |
| "name": "DeepSeek-R1 causal LM", | |
| "usage": "AutoTokenizer + AutoModelForCausalLM, batched prompts with left padding, causal LM loss on shifted labels", | |
| "architecture": "DeepseekV3ForCausalLM", | |
| "checkpoints": { | |
| "deepseek-ai/DeepSeek-R1": { | |
| "fixture": "huggingface-mic/DeepSeek-R1-mic" | |
| } | |
| }, | |
| "fixtures": { | |
| "huggingface-mic/DeepSeek-R1-mic": { | |
| "built_from": "deepseek-ai/DeepSeek-R1", | |
| "covers": [ | |
| "deepseek-ai/DeepSeek-R1" | |
| ], | |
| "why_separate": "The checkpoint is a 671B-parameter MoE model stored in FP8; the fixture is a randomly initialized slim model with the same architecture, tokenizer and vocabulary so the public API and loss path can be exercised on CPU.", | |
| "build": { | |
| "seed": 17, | |
| "model_class": "DeepseekV3ForCausalLM" | |
| } | |
| } | |
| }, | |
| "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": 50 | |
| }, | |
| "deepseek-ai/DeepSeek-R1": { | |
| "devices": [ | |
| "cuda" | |
| ], | |
| "gpus": 1, | |
| "gpu_memory_gb": 80, | |
| "memory_gb": 64, | |
| "download_mb": 700000, | |
| "variants": { | |
| "default": { | |
| "devices": [] | |
| }, | |
| "eager": { | |
| "devices": [] | |
| } | |
| } | |
| } | |
| } | |
| } | |
| }, | |
| "outputs": { | |
| "runtime": { | |
| "compare": "exact" | |
| }, | |
| "next_token_ids": { | |
| "compare": "exact" | |
| }, | |
| "probe_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" | |
| } | |
| } | |
| } | |