All issues

Search results

Track DeepSeek-V4.1-Flash on a four-Spark switchless cycle (Engram tables on NVMe, vLLM dsv41-feat + SparkRing NCCL) — WIP, PR to follow

Status: Open.

#260 In FujitsuPolycom/sparkring;

· sethforprivacy opened on Sep 10, 2026

Track nvidia/GLM-5.3-Flash-NVFP4 (official ModelOpt NVFP4, 2026-09-09) as a GLM-5.3 Flash target variant

Status: Open.

#257 In FujitsuPolycom/sparkring;

· sethforprivacy opened on Sep 10, 2026

GLM-5.3 Flash TP2 misidentifies decoded blue video as black

bugSomething isn't workingSomething isn't working

needs-triageAwaiting maintainer investigation and classificationAwaiting maintainer investigation and classification

Status: Open.

#229 In FujitsuPolycom/sparkring;

· FujitsuPolycom opened on Sep 6, 2026

GLM-5.3 TP4/DCP4: dynamic speculative depth (num_speculative_tokens_per_batch_size or adaptive_speculative_tokens_window) hangs CUDA-graph capture in pooled_indexer._project_head_weights on all ranks

Status: Open.

#221 In FujitsuPolycom/sparkring;

· sethforprivacy opened on Sep 5, 2026

[bug] Native MTP3: pinned sparse prefix retention may discard reachable cache states (SparkCache / TP4 DCP4)

bugSomething isn't workingSomething isn't working

Status: Open.

#220 In FujitsuPolycom/sparkring;

· FujitsuPolycom opened on Sep 5, 2026

[gb10-vllm-serving@827a8e8c] Named/required tool_choice contract violations escape as HTTP 200 with empty tool_calls

Status: Open.

#217 In FujitsuPolycom/sparkring;

· Saolence opened on Sep 5, 2026

GLM-5.3 TP4/DCP4: collective hang in cp_lse_ag_out_rs / mask_dcp_empty_shards_ under 4 concurrent 80k-300k conversations; rank 0 dies on the 300 s watchdog, liveness stays 200

Status: Open.

#215 In FujitsuPolycom/sparkring;

· sethforprivacy opened on Sep 4, 2026

GLM warmup_dflash.py samples at temperature 0 / thinking off: first temperature>0 request after readiness JITs _gumbel_sample_kernel (and fused_recurrent_kda_fwd_kernel on a 4K-prompt burst)

Status: Open.

#214 In FujitsuPolycom/sparkring;

· sethforprivacy opened on Sep 4, 2026

Perf: per-size LL/Simple allreduce tuner for the DeepSeek cycle runtime (AICAD Stage B port, NCCL 2.30.7)

Status: Open.

#193 In FujitsuPolycom/sparkring;

· Saolence opened on Sep 3, 2026

Perf: NCCL_BUFFSIZE=8M + IB timeout/retry/TOS in DeepSeek cycle env (AICAD T1aM4 evidence)

Status: Open.

#192 In FujitsuPolycom/sparkring;

· Saolence opened on Sep 3, 2026

Perf: A/B DSpark speculative depth k=7 vs k=5 for DeepSeek cycle (AICAD runs k=7)

Status: Open.

#191 In FujitsuPolycom/sparkring;

· Saolence opened on Sep 3, 2026

Proposal: enable PyTorch NCCL flight recorder for TP rank-loss / hang forensics (TORCH_FR_* on the /cache mount)

Status: Open.

#190 In FujitsuPolycom/sparkring;

· Saolence opened on Sep 3, 2026