New Model Request Reply

Workflow runs for New Model Request Reply

• Automatically reusing components

  skipped:
  New Model Request Reply#234:Issue#15248openedbyyiyixuxu
  Nov 11, 20261s
  More actions for workflow run 234

• [MiniMax-H3, Pika H3 Flex] ref2va fails on a reference soundtrack that is not at 32 kHz without torchaudio

  skipped:
  New Model Request Reply#233:Issue#15245openedbychristopher5106
  Nov 11, 202612s
  More actions for workflow run 233

• LLaDA2: EOS early stopping runs every remaining block to the step limit in batched generation

  skipped:
  New Model Request Reply#232:Issue#15242openedbyyyu-260
  Nov 11, 20261s
  More actions for workflow run 232

• Stable Audio 3 attention repeats RoPE cos / sin and q / k norm work on every call

  skipped:
  New Model Request Reply#231:Issue#15239openedbychsu-amd
  Nov 11, 20261s
  More actions for workflow run 231

• Docs: 17 dead external links (verified ×3) — upstream content removed, needs maintainer decisions

  skipped:
  New Model Request Reply#230:Issue#15233openedbyHisernberg
  Nov 10, 20262s
  More actions for workflow run 230

• How should ModularPipeline handle a component that failed to load?

  skipped:
  New Model Request Reply#229:Issue#15227openedbyyiyixuxu
  Nov 9, 20262s
  More actions for workflow run 229

• Wan transformer block makes float32 copies of hidden_states in its residual adds

  skipped:
  New Model Request Reply#228:Issue#15224openedbyklin2024
  Nov 9, 20267s
  More actions for workflow run 228

• FP32LayerNorm makes two full-size float32 copies on every call

  skipped:
  New Model Request Reply#227:Issue#15221openedbyklin2024
  Nov 9, 202611s
  More actions for workflow run 227

• RoPE helpers make float32 copies of q / k on every call

  skipped:
  New Model Request Reply#226:Issue#15220openedbyklin2024
  Nov 9, 202611s
  More actions for workflow run 226

• _tp_plan / _cp_plan: allow entries for modules that are not always present

  skipped:
  New Model Request Reply#225:Issue#15218openedbyyiyixuxu
  Nov 9, 20261s
  More actions for workflow run 225

• unload_textual_inversion(token) leaves the extra tokens of a multi-vector embedding behind, so it can't be loaded again

  skipped:
  New Model Request Reply#224:Issue#15211openedbysohailshk
  Nov 8, 20267s
  More actions for workflow run 224

• set_ip_adapter_scale(float) raises "Cannot assign 1 scale_configs to 2 IP-Adapter" when several IP-Adapters are loaded

  skipped:
  New Model Request Reply#223:Issue#15210openedbysohailshk
  Nov 8, 20262s
  More actions for workflow run 223

• [modular] Per-workflow input signature in the docstring of blocks with _workflow_map

  skipped:
  New Model Request Reply#222:Issue#15208openedbyyiyixuxu
  Nov 8, 20261s
  More actions for workflow run 222

• UniPCMultistepScheduler.set_timesteps(sigmas=...) raises TypeError for a list when use_flow_sigmas=True

  skipped:
  New Model Request Reply#221:Issue#15207openedbycharchit7
  Nov 7, 20262s
  More actions for workflow run 221

• device_map="balanced" drops all but the last CPU-overflow component, which then loads without offload hooks and breaks inference

  skipped:
  New Model Request Reply#220:Issue#15204openedbycharchit7
  Nov 7, 20261s
  More actions for workflow run 220

• FLEX attention silently ignores is_causal=True when no mask is supplied

  skipped:
  New Model Request Reply#219:Issue#15195openedbydavrens101
  Nov 6, 20261s
  More actions for workflow run 219

• HookRegistry.remove_hook drops the next hook's new_forward (breaks FirstBlockCache with group offloading)

  skipped:
  New Model Request Reply#218:Issue#15193openedbytheontho
  Nov 6, 20261s
  More actions for workflow run 218

• Per-token FlowMatch Euler steps use the opposite velocity sign

  skipped:
  New Model Request Reply#217:Issue#15185openedbydavrens101
  Nov 5, 20261s
  More actions for workflow run 217

• flash_varlen raises when requesting log-sum-exp

  skipped:
  New Model Request Reply#216:Issue#15182openedby0z5a
  Nov 4, 20261s
  More actions for workflow run 216

• flash_4_hub does not forward return_lse to its kernel

  skipped:
  New Model Request Reply#215:Issue#15180openedby0z5a
  Nov 4, 20269m 38s
  More actions for workflow run 215

• Gradient clipping is a no-op in train_text_to_image_lora.py and train_lcm_distill_lora_sdxl.py

  skipped:
  New Model Request Reply#214:Issue#15178openedbygorkemgul
  Nov 4, 20261s
  More actions for workflow run 214

• img2img: outputs paired with the wrong init image

  skipped:
  New Model Request Reply#213:Issue#15176openedbyArnav66692
  Nov 3, 20261s
  More actions for workflow run 213

• Cosmos 3 docs: FP8 snippet fails at the first W8A8 step without importing ModelOpt's FP8 GEMM backend

  skipped:
  New Model Request Reply#212:Issue#15170openedbypiotrm-nvidia
  Nov 2, 20261s
  More actions for workflow run 212

• Cosmos 3 FP8: wrapping pipe.transformer (e.g. torch.compile) silently disables the checkpoint's W8A16 step policy

  skipped:
  New Model Request Reply#211:Issue#15169openedbypiotrm-nvidia
  Nov 2, 20261s
  More actions for workflow run 211

• UniPC and DPM-Solver with use_flow_sigmas condition the model on integer timesteps while integrating with float sigmas

  skipped:
  New Model Request Reply#210:Issue#15167openedbyLiRunGuo
  Nov 2, 2026Skipped
  More actions for workflow run 210