Search results

PG timeouts never adapted if state dict is loaded

Status: Open.

Bug

#3152 In pytorch/torchtitan;

· hirschsn opened on Apr 29, 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

SFT eot_id and causality bugs

Status: Open.

Task

#3130 In pytorch/torchtitan;

· garrett361 opened on Apr 28, 2026

DeepSeek V3 16B crashes with 'tensor data not allocated' during backward with flex_attention + compile

Status: Open.

Bug

#3128 In pytorch/torchtitan;

· acisseJZhong opened on Apr 28, 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

replace vllm compile with trainer model compile?

Status: Open.

#3119 In pytorch/torchtitan;

· tianyu-l 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

Random weight init bug when using FSDP2 + CPU Offload

Status: Open.

Bug

#3106 In pytorch/torchtitan;

· man2machine opened on Apr 26, 2026