Yuchen Wang yuchenwang3

Yuchen (Ean) Wang — handwritten typing signature

Agentic post-training & ML systems.
M.S. CS @illinois · Research intern @Accio-Lab / @alibaba.

Research

Occamy-1.0

35B-A3B agent model for long-horizon tool use.
_{Core contributor · Post-training algorithms, infrastructure & data}

CineFlow

Dependency-driven parallel video generation.
_{Co-first author · 1.7–5.5× end-to-end speedup in the reported evaluation}

Dynamic Prefill

Adaptive batching and prompt packing for LLM serving.
_{Up to 20% lower TTFT on the reported traces}

Open-source contributions

11 merged · 1 adopted solution · 28 open · 17 projects

Highlights

HF Datasets	Preserve source shards during streaming shuffle so four workers can share 19 shards
mcore-bridge	Score packed QSA within each document; 2.70× faster in an 8K synthetic selector benchmark 2 more PRs in this repository Preserve low-precision rounding in gated residual mixing Bound PLE backward's extra workspace through chunked token reductions
vLLM	Remove full-weight copies during MoE loading; conversion peak 7.88 → 3.94 GiB in the exact-shape TP2 benchmark 1 more PR in this repository Clarify Qwen3 parser boundary tokens in custom grammars
NeMo RL	Bypass driver tensor materialization in distillation; 4.4–5.3× faster transfers in a controlled Ray benchmark 3 more PRs in this repository Render evaluation prompts as complete conversations Unify worker selection through configuration Sanitize non-finite async log probabilities
Megatron-LM	Fuse GDN Q/K normalization to remove an extra backward activation buffer Enable selective Mamba mixer recompute to save activation memory without full-layer recomputation 5 more PRs in this repository Preserve native Adam step counters across checkpoint restoration Reuse packed chunkwise CP metadata across GDN and KDA layers Route GDN input projections to Adam Exclude GDN input projections from global clipping Skip gradient clipping for Muon
FlashAttention	Prevent redundant backward-kernel recompilation with stable cache keys, without CPU–GPU sync _{Solution adopted by the PR author}
FlashInfer	Restore K/V calibration in FP8 KV prefill, correcting silently mis-scaled attention outputs
SGLang	Publish Mamba cache updates before dependent batches capture stale KV mappings 3 more PRs in this repository Reject developer messages silently dropped by templates Explain cold MXFP4 JIT startup Honor weight-check exclusions during reset
Hermes Agent	Resolve local schema references so nested tool arguments reach handlers as objects, not JSON strings 3 more PRs in this repository Control partial-stream continuation for batch evaluation Clarify API retry budgets and streaming defaults Make SSH reconnects race-safe