Issues

• Full GLM-5.3 (glm_moe_dsa) on three DGX Sparks: a native CUDA family (step 5 of #593)

  enhancement
  ashhart/TensorFold#3063 · Filter by author bjornmonnens opened on Nov 12, 2026

• glm cuda: support explicitly paged FP8 KV consistently

  ashhart/TensorFold#3062 · Filter by author gilp456 opened on Nov 12, 2026
  6

• Nemotron on HIP: read oQ mixed-width MLX checkpoints (4, 5, 6 and 8 bits, a width per tensor)

  ashhart/TensorFold#3061 · Filter by author gilby opened on Nov 12, 2026

• XPU (B70) Qwen3.8-27B: serve logprobs and top_logprobs

  ashhart/TensorFold#3050 · Filter by author Mirrdhyn opened on Nov 12, 2026
  1

• Gemma 4 on Metal: a new prompt freezes every other stream's decode for its whole prefill (no prefill stepping)

  ashhart/TensorFold#3021 · Filter by author ashhart opened on Nov 12, 2026

• GLM CUDA: honor full-prefix repetition_penalty from generation defaults and API requests

  ashhart/TensorFold#3020 · Filter by author gilp456 opened on Nov 12, 2026
  3

• Qwen3.8-27B + DFlash2 on Metal: one greedy stream's copy-heavy output is ~70% slower since the 16-row lane loop became the default (tokens exact)

  ashhart/TensorFold#3013 · Filter by author ashhart opened on Nov 12, 2026

• Flash Next on CUDA: agent turns re-prefill the full history — opt-in message-start cuts or a warm turn (as #1334/#2396/#2824 did for other families)

  ashhart/TensorFold#3009 · Filter by author arnon2020 opened on Nov 12, 2026
  8

• /v1/models: no input-modality discovery — clients can't tell whether a model accepts image/video parts

  ashhart/TensorFold#3004 · Filter by author WebReflection opened on Nov 12, 2026
  2

• HIP: on gfx1151 the matrix-core prompt tiles are up to 2.9x slower on short passes when a weight row is 1,024, 2,048 or 4,096 bytes (K 2048, 4096, 8192 at 4-bit); 64 bytes of padding a row removes it

  ashhart/TensorFold#3001 · Filter by author AbrahamCete opened on Nov 12, 2026
  4

• CUDA on unified memory: what the reserve, the KV pool and the prompt cache each decide, and at which scope

  ashhart/TensorFold#3000 · Filter by author Bizuayeu opened on Nov 12, 2026
  2

• Native SAM3/SAM3.1 segmentation and tracking on Metal

  ashhart/TensorFold#2991 · Filter by author CerebralCoding opened on Nov 12, 2026
  2

• H3 Metal: read checkpoint shards ahead of their first mapped access

  ashhart/TensorFold#2988 · Filter by author plotarmordev opened on Nov 12, 2026
  3

• GLM CUDA: a named TF_KV_SHARE is planned after every window the room holds, so at TP=3 on a lighter checkpoint it keeps 8 of 32 states

  ashhart/TensorFold#2987 · Filter by author Bizuayeu opened on Nov 12, 2026
  1

• DeepSeek-V4.1 on Metal: decode and prompt speed on M3 Ultra and M5 Ultra, token-exact (gate and order for the drafts)

  ashhart/TensorFold#2984 · Filter by author gilby opened on Nov 12, 2026
  4

• Flash Next on CUDA: flashnext-rank timeout stops both ranks inside NCCL initialization

  ashhart/TensorFold#2975 · Filter by author swbiggart opened on Nov 12, 2026
  3

• Flash Next on CUDA: TF_FLASHNEXT_DECODE_GRAPH=1 at --tp 2 fails the first plain round

  ashhart/TensorFold#2974 · Filter by author swbiggart opened on Nov 12, 2026
  3

• Flash Next on CUDA: a --tp 2 start fails when one rank loads over 30 s later

  ashhart/TensorFold#2970 · Filter by author swbiggart opened on Nov 12, 2026
  2

• Nemotron on Metal without an MTP head: concurrent streams share one round

  ashhart/TensorFold#2968 · Filter by author ashhart opened on Nov 12, 2026

• Nemotron CUDA: a failed image encode leaves its rows to splice into the next request

  ashhart/TensorFold#2943 · Filter by author ashhart opened on Nov 11, 2026
  5

• Qwen 3.5 CUDA: a prompt chunk nothing reads stops its last layer at the keys

  ashhart/TensorFold#2941 · Filter by author ashhart opened on Nov 11, 2026
  2

• qwen27 Metal: K and V read in place from the mixed projection, no unstack copies

  ashhart/TensorFold#2940 · Filter by author ashhart opened on Nov 11, 2026

• qwen27 Metal (M5): out_proj group sums written by the GDN post and attention merge kernels

  ashhart/TensorFold#2931 · Filter by author ashhart opened on Nov 11, 2026

• qwen27 Metal: greedy picks ride the verify's last command buffer instead of their own

  ashhart/TensorFold#2930 · Filter by author ashhart opened on Nov 11, 2026

• HIP: on gfx1151 a plain prompt pass of 1 or 2 rows runs on kp0; on kp3 a resumed first token is 18-24% sooner on Qwen3.6-35B-A3B and 48-64% on Qwen3.8-27B, same bytes

  ashhart/TensorFold#2923 · Filter by author AbrahamCete opened on Nov 11, 2026
  5