RED-SNOW 5.3 FLASH DFlash2 Stage 2

Private, target-specific DFlash2 draft checkpoint for the RED-SNOW 5.3 FLASH model family. This is the consolidated serving artifact from the promoted stage-2 checkpoint (epoch_1_step_768). It is a speculative-decoding draft, not a standalone chat model.

Artifact

  • Architecture: DFlash2DraftModel
  • Precision: BF16
  • Draft layers: 5
  • Target feature layers: [5, 14, 24, 33, 42]
  • DFlash block size: 8
  • Maximum position embeddings: 1,048,576
  • Weight shard: model-00001-of-00001.safetensors

Download

hf download Blackfrost-AI/RED-SNOW-5.3-FLASH-DFlash2-STAGE2 \
  --local-dir /workspace/models/RED-SNOW-5.3-FLASH-DFlash2-STAGE2 \
  --max-workers 4

SGLang integration

Point SGLang at this directory with the DFlash speculative-decoding options:

--speculative-algorithm DFLASH
--speculative-draft-model-path /path/to/RED-SNOW-5.3-FLASH-DFlash2-STAGE2
--speculative-dflash-block-size 8
--speculative-num-draft-tokens 8
--speculative-draft-attention-backend fa4
--speculative-draft-kv-cache-dtype bfloat16
--speculative-draft-model-quantization unquant

Use it only with a tokenizer-compatible RED-SNOW 5.3 FLASH target. Keep the target model's normal reasoning parser, tool parser, chat template, and tensor parallel settings.

Integrity

SHA-256 of model-00001-of-00001.safetensors:

6be9ccb43cc6f641739c4dc0ecc93aa55fa6af788d131729c564e975cdeb0148
Downloads last month
19
Safetensors
Model size
1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support