GitLab

[BUGFIX] 修复 Qwen3-MoE Attention 中 fused RMS RoPE 的 epsilon 参数顺序错误

See merge request dcutoolkit/deeplearing/vllm!539

feat: Support rms+quant fusion in minimax_m2 series model.

See merge request dcutoolkit/deeplearing/vllm!538

use tunning w4a16 moe

See merge request dcutoolkit/deeplearing/vllm!535

feat(v1 attention): 为 ROCm FlashAttention 接入 unified kv layout，并打通 mm_prefix、qq_bias 与 use_alibi_sqrt 透传

See merge request dcutoolkit/deeplearing/vllm!526

feat(v1 attention): 为 ROCm FlashAttention 接入 unified kv layout，并打通 mm_prefix、qq_bias 与 use_alibi_sqrt 透传
在 ROCm FlashAttention 后端增加 unified KV layout 选择逻辑
接入 unified varlen kernel 调用路径
在 FlashAttention metadata 中补充 mm_prefix_range 与 qq_bias 透传

The gfx928 architecture force to set VLLM_W8A8_BACKEND == 1

See merge request dcutoolkit/deeplearing/vllm!533

glm5 融合算子优化

See merge request dcutoolkit/deeplearing/vllm!534

支持kvacache fp8_e4m3的RMS_ROPE_CONCAT

See merge request dcutoolkit/deeplearing/vllm!531

处理VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD导致的awq推理bug问题

See merge request dcutoolkit/deeplearing/vllm!528

fix(moe): 补齐非Marlin量化路径 shared_output/routed_scaling_factor 透传

See merge request dcutoolkit/deeplearing/vllm!529

支持kvacache fp8_e4m3/fp8_e5m2的RMS_ROPE_CONCAT

Gy v015 1dmrope

See merge request dcutoolkit/deeplearing/vllm!530

支持kvacache fp8_e4m3/fp8_e5m2的RMS_ROPE_CONCAT

删除1d_mrope

See merge request dcutoolkit/deeplearing/vllm!525

关闭sparse_mla的num_head到64/128的pad，以及添加控制fp8_use_mixed_batch模式的环境变量控制，FP8_USE_MI...

See merge request dcutoolkit/deeplearing/vllm!524

# Conflicts:
#   vllm/model_executor/layers/sparse_attn_indexer.py
#   vllm/v1/attention/backends/mla/flashmla_sparse.py

feat:flash_mla，q去掉pad

See merge request dcutoolkit/deeplearing/vllm!522

关闭sparse_mla的num_head到64/128的pad，以及添加控制fp8_use_mixed_batch模式的环境变量控制，FP8_USE_MIXED_BATCH，默认为false，为分离模式