All pull requests

Pull requests

• Gemma4 vision: Instance level refactoring norm and attention

  #1316·nikheal2 opened Oct 19, 2026Contributor·4 of 5 tasks·

  ·
  1

  2

• ci: report Spyre libraries loaded from the image instead of the lock

  #1315·joerunde opened Oct 19, 2026Collaborator·5 tasks·

  ·
  2

• chore(deps): bump torch-spyre to 54d9cc2

  sentinel-alert
  #1306·yannicks1 opened Oct 19, 2026Collaborator·10 of 13 tasks·

  ·
  8

• perf(gemma4): reduce the dense-MLP and MoE outputs with one all-reduce

  #1305·jvlunteren opened Oct 19, 2026Collaborator·4 of 5 tasks·

  ·
  1

• fix(moe): read the gathered expert index through a reduction, not a slice of the staggered float32

  #1304·yoheiueda opened Oct 19, 2026·3 of 5 tasks·

  ·
  1

• fix: shim head_dim on Attention classes inherited from a base class's module

  #1303·biggdawg320 opened Oct 19, 2026·

  ·
  1

  1

• perf(attn): group multi-bucket batched decode

  #1301·jvlunteren opened Oct 17, 2026Collaborator·4 of 5 tasks·

  ·
  1

• Add direct-to-device SpyreModelLoader (--load-format spyre)

  sentinel-alert
  #1294·amritahs-ibm opened Oct 16, 2026·5 tasks·

  ·
  6

• [WiP] Minimize d2h transfers reworks

  #1286·yannicks1 opened Oct 15, 2026Collaborator·5 tasks·

  ·
  2

• perf(logits): gather TP logits shards on the CPU behind SPYRE_LOGITS_CPU_GATHER

  #1283·jvlunteren opened Oct 15, 2026Collaborator·5 tasks done·

  ·
  1

• Add CLIP model benchmark definition

  #1280·dmholtz opened Oct 15, 2026Contributor·2 of 8 tasks·

  ·
  2

• fix(conv): skip input staging in SpyreConv2d

  #1279·gkumbhat opened Oct 15, 2026Collaborator·5 tasks·

  ·
  1

• [MRV2] Platform glue: device shims, Triton gate, opt-in switch

  #1277·joerunde opened Oct 14, 2026Collaborator·5 tasks done·

  ·
  40

• bench: add multimodal serve benchmark config

  ready
  #1276·gkumbhat opened Oct 14, 2026Collaborator·3 of 5 tasks·

  ·
  7

• feat(tests): op-test YAML generated through spyre-inference

  #1267·gokulsp05 opened Oct 14, 2026·

  ·
  2

• Cherry pick skill

  #1266·bohnstingl opened Oct 14, 2026Collaborator·4 of 5 tasks·

  ·
  13

• feat(multimodal): compile and warm the Pixtral vision encoder per bucket

  phase-2readysentinel-alert
  #1261·gkumbhat opened Oct 13, 2026Collaborator·Queued·5 tasks
  2

  30

• feat(ci): run torch-spyre model-ops tests over tests/model_ops YAMLs (make test-model-ops)

  sentinel-alert
  #1255·anubhavjana opened Oct 13, 2026Contributor·

  ·
  6

• ci: run the model-ops op tests over tests/model_ops/*.yaml

  sentinel-alert
  #1253·HarikrishnanBalagopal opened Oct 13, 2026Contributor·

  ·
  5

• Fix Spyre device selection from PCIe addresses

  phase-2
  #1250·AnishPahilajani opened Oct 13, 2026Contributor·5 tasks·

  ·
  2

• Perf/bge long context body tiers

  #1226·GOavi101 opened Oct 12, 2026Contributor·4 of 5 tasks·

  ·
  1

  19

• ci(ingest): install the ingest library once, pin setup-uv, raise timeout

  sentinel-alert
  #1224·ashokponkumar opened Oct 12, 2026Contributor·

  ·
  2

• multimodal/granite4_vision: on-device block-mean for integer downsample ratios

  #1217·ktuvw opened Oct 10, 2026Contributor·4 of 5 tasks·

  ·
  1

  1

• multimodal/granite4_vision: shrink _ds_buffer staging tensor to actual range

  #1213·ktuvw opened Oct 9, 2026Contributor·4 of 5 tasks·

  ·
  1

  1

• feat(attention): add a denser KV ladder for large batched decode batches

  phase-2
  #1209·jvlunteren opened Oct 9, 2026Collaborator·5 tasks done·

  ·
  20