SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD
Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orches
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
Every edge carries a method, confidence, and the source snippet that justified it — so bad links are debuggable.
- FuzzySimilar title/name (fuzzy) · 87%deepseek-ai/DeepSeek-R1 →
“Fuzzy title match (0.94): “SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 ” ≈ “deepseek-ai/DeepSeek-R1””
- FuzzySimilar title/name (fuzzy) · 87%deepseek-ai/DeepSeek-OCR →
“Fuzzy title match (0.94): “SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 ” ≈ “deepseek-ai/DeepSeek-OCR””
- FuzzySimilar title/name (fuzzy) · 87%deepseek-ai/DeepSeek-V4-Pro →
“Fuzzy title match (0.94): “SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 ” ≈ “deepseek-ai/DeepSeek-V4-Pro””
- FuzzySimilar title/name (fuzzy) · 84%deepseek-ai/DeepSeek-V3 →
“Fuzzy title match (0.92): “SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 ” ≈ “deepseek-ai/DeepSeek-V3””
- FuzzySimilar title/name (fuzzy) · 66%deepseek-ai/DeepSeek-V3-0324 →
“Fuzzy title match (0.78): “SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 ” ≈ “deepseek-ai/DeepSeek-V3-0324””
- FuzzyOverlapping authors or contributors · 78%sgl-project/sglang →
“Shared author/contributor keys: luo, zhou”
- FuzzySimilar title/name (fuzzy) · 66%esengine/DeepSeek-Reasonix →
“Fuzzy title match (0.78): “SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 ” ≈ “esengine/DeepSeek-Reasonix””
- FuzzyOverlapping authors or contributors · 62%ray-project/ray →
“Shared author/contributor keys: wang”