Issues

• AsyncPutManager raises an async put failure twice: once from wait_for_buffer, again from the next check_last_error

  lightseekorg/TorchSpec#218 · Filter by author alpaylan opened on Nov 3, 2026

• K3Renderer.get_assistant_token_ids() does not match named assistant turns, so the dynamic loss mask drops them

  lightseekorg/TorchSpec#217 · Filter by author alpaylan opened on Nov 2, 2026

• MiniMaxParser.format splits assistant content at "</think>" differently from the MiniMax-M2.5 chat template

  lightseekorg/TorchSpec#216 · Filter by author alpaylan opened on Nov 2, 2026

• DFlash attention ignores rope_scaling: drafts train with unscaled RoPE while the config declares scaling

  lightseekorg/TorchSpec#215 · Filter by author alpaylan opened on Nov 2, 2026
  1

• DFlashConfig / DSparkConfig raise KeyError on a Llama-3.1-style rope_theta + llama3 rope_scaling (transformers 5.x)

  lightseekorg/TorchSpec#214 · Filter by author alpaylan opened on Nov 2, 2026

• DFlashConfig / DSparkConfig crash with AttributeError when given a YaRN rope_scaling (transformers 5.x)

  lightseekorg/TorchSpec#213 · Filter by author alpaylan opened on Nov 2, 2026

• KimiK25Parser writes dict tool-call arguments as a Python repr ({'q': 'a'}) instead of JSON

  lightseekorg/TorchSpec#212 · Filter by author alpaylan opened on Nov 2, 2026

• KimiK25Parser renders a message with content: null as the text "None" and trains on it

  lightseekorg/TorchSpec#211 · Filter by author alpaylan opened on Nov 2, 2026

• EAGLE3 -> EAGLE-4 attn_norm warm start fails with shape_mismatch when stacked q/k/v is narrower than hidden_size

  lightseekorg/TorchSpec#210 · Filter by author alpaylan opened on Nov 2, 2026

• generate_vocab_mapping crashes with ValueError when a dataset row has exactly one token

  lightseekorg/TorchSpec#209 · Filter by author alpaylan opened on Nov 2, 2026

• build_rotary_embedding ignores rope_theta for linear, dynamic and mrope RoPE scaling (drafts use base 10000)

  lightseekorg/TorchSpec#208 · Filter by author alpaylan opened on Nov 2, 2026

• build_rotary_embedding crashes with TypeError on a YaRN rope_scaling that omits beta_fast/beta_slow

  lightseekorg/TorchSpec#207 · Filter by author alpaylan opened on Nov 2, 2026

• compute_assistant_loss_mask(last_turn_only=True) masks an earlier assistant turn when the last one is empty

  lightseekorg/TorchSpec#206 · Filter by author alpaylan opened on Nov 2, 2026

• SmgRenderer with minimax-m2 ends the assistant span inside the content when the content ends in "[e~"

  lightseekorg/TorchSpec#205 · Filter by author alpaylan opened on Nov 2, 2026

• SMG binding silently rounds integers outside the i64/u64 range in tool-call arguments to floats

  lightseekorg/TorchSpec#204 · Filter by author alpaylan opened on Nov 2, 2026

• Development Roadmap (2026 Q4)

  high priority
  lightseekorg/TorchSpec#194 · Filter by author Dogacel opened on Sep 10, 2026