diff --git a/examples/configs/offline/colocated/qwen3.5-4b-dflash-offline-amd.yaml b/examples/configs/offline/colocated/qwen3.5-4b-dflash-offline-amd.yaml index 2d909b424..70bedc43b 100644 --- a/examples/configs/offline/colocated/qwen3.5-4b-dflash-offline-amd.yaml +++ b/examples/configs/offline/colocated/qwen3.5-4b-dflash-offline-amd.yaml @@ -9,6 +9,9 @@ model: trust_remote_code: true embedding_key: "model.language_model.embed_tokens.weight" torch_dtype: "bfloat16" + # Liger fused RMSNorm/SwiGLU: validated on MI355X (+10% tokens/s, -15% peak + # mem, loss parity). Requires the flex-backend fix (torch<2.11) to run. + use_liger_kernel: true data: hidden_states_path: "./cache/hidden_states/qwen3.5-4b-dflash-sharegpt" max_length: 2048 diff --git a/examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-amd.yaml b/examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-amd.yaml index af3bac849..a37cd4d61 100644 --- a/examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-amd.yaml +++ b/examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-amd.yaml @@ -10,6 +10,9 @@ model: trust_remote_code: true embedding_key: "model.language_model.embed_tokens.weight" torch_dtype: "bfloat16" + # Liger fused RMSNorm/SwiGLU: validated on MI355X (+10% tokens/s, -15% peak + # mem, loss parity). Requires the flex-backend fix (torch<2.11) to run. + use_liger_kernel: true data: train_data_path: "./cache/dataset/sharegpt_train.jsonl" max_length: 2048