All issues

Search results

DSpark: confidence-head BCE is weighted with the D-PACE drafter loss weights when --per-position-loss-weight dpace is set

Status: Open.

#1171 In vllm-project/speculators;

· pipa83165 opened on Sep 27, 2026

A DFlash2 speculative decoding draft model for gemma-4-E2B-it (167 --> 626 tokens/s; up to 3.74x speedup on Math)

Status: Open.

#1169 In vllm-project/speculators;

· rasyosef opened on Sep 26, 2026

prepare_data silently drops remaining conversations when the render server dies mid-run

Status: Open.

#1159 In vllm-project/speculators;

· ra-srivid opened on Sep 24, 2026

[Question] Which vLLM Ascend version is compatible with Speculators 0.8.0

Status: Open.

#1157 In vllm-project/speculators;

· PixelFlat opened on Sep 23, 2026

[Question] Why choose guidellm instead of vllm bench for speculator evaluation?

Status: Open.

#1151 In vllm-project/speculators;

· zhujiem opened on Sep 19, 2026

[Question]: DFlash2 multimodal acceptance gap with mixed text+MM training — text deploy ≥ train val, but ALLaVA deploy −10%

Status: Open.

#1142 In vllm-project/speculators;

· Brilitan opened on Sep 18, 2026

DFlash2 asserts an exact layer class, making the documented _make_decoder_layer override unusable

Status: Open.

#1137 In vllm-project/speculators;

· Hyungkeun-Park opened on Sep 18, 2026

[RFC]: DSpark draft models for Llama-3.2-3B-Instruct & gemma-4-E2B-it (5.5 accept length on math_reasoning)

RFC

Status: Open.

#1135 In vllm-project/speculators;

· rasyosef opened on Sep 17, 2026

[RFC]: Pretrain DFlash-family drafts on raw text before distillation

Status: Open.

#1127 In vllm-project/speculators;

· reed-meyerson opened on Sep 15, 2026

[RFC]: Training-run integrity: skip steps on non-finite loss, log gradient norm, always persist the final checkpoint

Status: Open.

#1120 In vllm-project/speculators;

· Leslie360 opened on Sep 15, 2026

[Bug]: Eagle3 conversion writes randomly initialized weights when a checkpoint key does not match, instead of failing

Status: Open.

#1114 In vllm-project/speculators;

· yupengtang opened on Sep 12, 2026

[Bug]: Per-engine counter decreases can be hidden by aggregate acceptance deltas

Status: Open.

#1106 In vllm-project/speculators;

· LOGO127 opened on Sep 9, 2026