Chunk 44.1

The B200 generation run completed successfully, producing 902,087 completions with full Qwen3.6-27B thinking traces (1.64B output tokens, 7.25 GB in S3). Analysis of the generated data confirmed that tool-calling prompts (12.5% of the dataset) produced proper JSON function calls with reasoning traces, though some degenerate `<tool_call>` loops appeared when the model expected tool execution feedback that never came. Multi-turn conversations (8.4%) had their assistant turns stripped as designed, with the model seeing only user messages. A critical architectural decision was made: offline hidden state extraction would require ~90 TB of storage (5 layers × 5120 hidden × BF16 × 2000 avg tokens × 902K samples), making it completely impractical. The team pivoted to an **online training approach** where hidden states are extracted on-the-fly during the target model forward pass and fed directly to the drafter, eliminating storage entirely. The 2× data-parallel architecture was designed: GPUs 0 and 1 each run a frozen copy of Qwen3.6-27B with hook-based extraction, transferring hidden states over PCIe Gen5 to GPUs 2 and 3 which hold the drafter and optimizer, with manual gradient synchronization between the two streams. Three scripts were implemented and syntax-validated: `dflash_model.py` (standalone DFlash drafter with flex attention, anchor selection, and block-diffusion loss), `tokenize_completions.py` (Phase 1: downloads 1,805 JSONL files from S3, applies Qwen3.6 chat template with thinking tokens, generates loss masks), and `train_dflash_online.py` (Phase 2+3: online extraction + training with 2× DP, checkpointing, S3 upload). Tokenization ran locally with 128 workers, completing 902,087 samples in 6.5 minutes — producing 1.87B tokens (87.5% loss tokens), a 5.75× improvement over the old prompt-only dataset. The tokenized data (47 Arrow shards) was uploaded to S3, and PROGRESS.md was updated with the complete pipeline state and detailed task list for provisioning the 4× PRO 6000 Blackwell instance.

From 90 Terabytes to Zero: The Online Training Pivot That Saved the DFlash Pipeline 2677 words

Message Articles