Repository Issue Activity (beta)

#2151 GeneralMoEWrapper.load_weights() cpu_save path reads nonexistent dict keys, guaranteed KeyError

#2035 使用L20*8部署deepseek_v4_flash失败

#1608 🚨FAQs | 常见问题🚨

#2127 DeepSeek-V4-Flash-0731 MTP failed to start on avx2 + sm89

#2149 kt-sft: kt-managed expert LoRA absent from every parameter count in logs (patch included)

#2148 kt-sft: fused-expert models fail to load from unsharded checkpoints (is_fused decided only from shard index)

#2147 kt-sft: kt_use_lora_experts + FSDP2 crashes at save and silently discards all trained LoRA-Expert weights

#2146 kt-sft: 'trainable params' underreports kt-managed fused expert LoRA by ~60x (10.6M shown, 640M trained)

#2145 kt-sft + FSDP2: DTensor _foreach_mul_ crash when kt expert-LoRA params mix with DTensor lists in grad clipping

#2144 kt-sft: _compute_routing branch ordering sends Linear-gate models down the TopKRouter path (IndexError)

#2150 kt-sft: chunked_prefill_size sized from cutoff_len, never scaled by batch — any per_device_train_batch_size>1 hard-errors

#2099 Out of memory error using Ktransformers with Deepseek v4 flash

#2106 Error with Deepskeep 4 flash with trtllm_fp4_block_scale_moe

#2110 convert_gpu_weights: AttributeError: module 'torch.accelerator' has no attribute 'get_memory_info'

#2128 AMXInt4 not supported in 2 numa machine with "Segmentation fault"

#2139 When will kvcache‑ai/sglang be synced with sgl‑project/sglang

#2113 MiniMax M2.7: AttributeError: 'LlamafileMoEWrapper' object has no attribute 'submit_write_weight_scale_to_buffer'

#2118 Support DSpark speculative decoding for DeepSeek-V4-Flash-0731 in sglang-kt

#2109 kimi k3啥时候支持？

#2115 ROCm Support for Dual AMD Radeon Pro V620 (GLM-5.2)

#2108 GLM5.2 int8: 'AMXMoEWrapper' object has no attribute 'submit_write_weight_scale_to_buffer'

#2107 RuntimeError: Only Tensors of floating point and complex dtype can require gradients

#2083 GLM5.2 on RTX 6000 Pro ( sm120 ) * Update

#2100 [Bug] 无法正确识别模型文件

#2040 部署Qwen3.6-35B-A3B-FP8出现的问题