vllm
[Feature][ROCm][Quantization] Requantize DeepSeek-V4.1 MXFP4 experts to block FP8 at load time
#69240
Open

Commits
  • [Feature][ROCm][Quantization] Requantize MXFP4 MoE weights to block FP8 online
    yuzhouo7 committed 23 hours ago
  • [ROCm][Quantization] Guard block FP8 requantization on clamped SiLU
    yuzhouo7 committed 13 hours ago
  • [Spec Decode] Apply online quantization targets to embedded DeepSeek DSpark
    yuzhouo7 committed 13 hours ago
  • [ROCm] Match DeepSeek V4.1 TP8 online block-FP8 to checkpoint inference
    yuzhouo7 committed 6 hours ago
Loading