Repository navigation

[OpenVINO] Support MiniCPM-OCR-1.0-DSpark: DSpark drafts with a pruned vocabulary - #2099

Open

imargulis wants to merge 18 commits into

huggingface:mainfrom

imargulis:dflash2-pruned-vocab

Open

[OpenVINO] Support MiniCPM-OCR-1.0-DSpark: DSpark drafts with a pruned vocabulary#2099
imargulis wants to merge 18 commits into
huggingface:mainfrom
imargulis:dflash2-pruned-vocab

Commits on Nov 7, 2026
- Keep the whole past in the DFlash draft cache with sliding layers
  
  imargulis
  and
  cursoragent
  committed
  a596215
Commits on Nov 9, 2026
- Base DFlash-2 on the per-row draft and export its selector as a Markov head
  
  committed
  6a95768
- Look up the draft architecture once
  
  imargulis
  and
  cursoragent
  committed
  d66ee4f
- Write DFlash block rows with index_put_ so the trace keeps their count dynamic
  
  imargulis
  and
  cursoragent
  committed
  bddb8c3
- Infer the DFlash export task for local drafts and test INT4 drafts on all layers
  
  imargulis
  and
  cursoragent
  committed
  03fcf7f
- Let the DFlash draft build its own cache in the PagedAttention reference
  
  imargulis
  and
  cursoragent
  committed
  537658f
- Derive the DFlash Markov head spec once per draft export config
  
  imargulis
  and
  cursoragent
  committed
  c975fad
- Mark the DFlash Markov head export config with a flag instead of its Top-K
  
  imargulis
  and
  cursoragent
  committed
  c6a5565
- Reword the Markov head part of the Eagle-3 dummy generator docstring
  
  imargulis
  and
  cursoragent
  committed
  ae02558
- Build the DFlash Markov head inputs from the shared ones
  
  imargulis
  and
  cursoragent
  committed
  407fec5
- Say what a masked DFlash-2 convolution tap reaches since block-row-only compute
  
  imargulis
  and
  cursoragent
  committed
  202ea0a
- Drop DFlash tests that other tests already cover
  
  imargulis
  and
  cursoragent
  committed
  d6fc424
- Drop the Markov head interface tag
  
  imargulis
  and
  cursoragent
  committed
  17168c9
- Leave Muse-Glimmer drafts out of the DFlash config validation
  
  imargulis
  and
  cursoragent
  committed
  e7bf1ae
- Annotate the hidden states of MiniCPM-OCR targets
  
  imargulis
  and
  cursoragent
  committed
  ee1d122
- Export DSpark drafts with a pruned draft vocabulary
  
  imargulis
  and
  cursoragent
  committed
  cad027e
- Pair the tiny MiniCPM-OCR DSpark draft with its target in the tests
  
  imargulis
  and
  cursoragent
  committed
  6b9cee4
- Compile the pruned vocabulary DSpark test models in f32
  
  imargulis
  and
  cursoragent
  committed
  c5221a2

Commits

Commits on Nov 7, 2026

Commits on Nov 9, 2026