Search results

[rl] FAv2 + num_splits=auto + paged kv produced NaN during inference

Status: Open.

Bug

#3235 In pytorch/torchtitan;

· wwwjn opened on May 6, 2026

[Quantization] Re-consolidate MXFP8 and FP8 linear converters

Status: Open.

Task

#3219 In pytorch/torchtitan;

· aditvenk opened on May 4, 2026

[GraphTrainer] Loss region FX nodes missing stacktrace and module_fqn on some backward nodes

Status: Open.

#3208 In pytorch/torchtitan;

· SherlockNoMad opened on May 4, 2026

Qwen3 moe: EP+DP init weight error

Status: Open.

Bug

#3205 In pytorch/torchtitan;

· pretidav opened on May 4, 2026

[RL] Support Cudagraph on Trainer

Status: Open.

#3175 In pytorch/torchtitan;

· Lucaskabela opened on Apr 30, 2026

[Quantization] MXFP8LinearConverter should offer filter_fqns instead of fqns

module: float8

questionFurther information is requestedFurther information is requested

Status: Open.

#3150 In pytorch/torchtitan;

· aditvenk opened on Apr 29, 2026

[SimpleFSDP] SimpleFSDP silently ignores enable_fsdp_float8_all_gather=True

Status: Open.

#3123 In pytorch/torchtitan;

· aditvenk opened on Apr 28, 2026

torch.compile + DTensor: backward crash on TP-parallelized sub-module with a param-shadowing pre-hook

Status: Open.

#3122 In pytorch/torchtitan;

· mreso opened on Apr 27, 2026

[graph_trainer] Chunked CE loss doesn't work in graph trainer

Status: Open.

#3117 In pytorch/torchtitan;

· yiming0416 opened on Apr 27, 2026

PP backward crashes on non-tensor forward kwargs ('bool' object has no attribute 'requires_grad')

Status: Open.

Feature

#3112 In pytorch/torchtitan;

· wwwjn opened on Apr 27, 2026

[rl] Fixed length train batch

Status: Open.

#3110 In pytorch/torchtitan;

· joecummings opened on Apr 27, 2026

Bug: FP8 + TP≥2 causes NaN loss from step 3 (use_local_output=False incompatible with torchao amax)

Status: Open.

Bug

#3109 In pytorch/torchtitan;

· zzhaobh opened on Apr 27, 2026