Add native Qwen MTP speculation

This commit is contained in:
Georg Bauer
2026-09-03 22:38:52 +02:00
parent 714b39f7f3
commit bf82df77cb
14 changed files with 1302 additions and 78 deletions

View File

@@ -2492,15 +2492,15 @@ mtp.safetensors mtp.layers.0.mlp.shared_expert.up_proj.weight U32 640x320 4 64 a
mtp.safetensors mtp.layers.0.mlp.shared_expert_gate.biases BF16 1x40 4 64 affine 140045196 140045276
mtp.safetensors mtp.layers.0.mlp.shared_expert_gate.scales BF16 1x40 4 64 affine 140045276 140045356
mtp.safetensors mtp.layers.0.mlp.shared_expert_gate.weight U32 1x320 4 64 affine 620915756 620917036
mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.biases BF16 512x2560x20 2 64 affine 140966956 193395756
mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.scales BF16 512x2560x20 2 64 affine 641468716 693897516
mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.weight U32 512x2560x80 2 64 affine 1253145132 1672575532
mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.biases BF16 512x640x80 2 64 affine 693897772 746326572
mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.scales BF16 512x640x80 2 64 affine 20277900 72706700
mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.weight U32 512x640x320 2 64 affine 760314412 1179744812
mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.biases BF16 512x640x80 2 64 affine 87616396 140045196
mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.scales BF16 512x640x80 2 64 affine 1192852012 1245280812
mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.weight U32 512x640x320 2 64 affine 194808876 614239276
mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.biases BF16 512x2560x20 4 32 affine 140966956 193395756
mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.scales BF16 512x2560x20 4 32 affine 641468716 693897516
mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.weight U32 512x2560x80 4 32 affine 1253145132 1672575532
mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.biases BF16 512x640x80 4 32 affine 693897772 746326572
mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.scales BF16 512x640x80 4 32 affine 20277900 72706700
mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.weight U32 512x640x320 4 32 affine 760314412 1179744812
mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.biases BF16 512x640x80 4 32 affine 87616396 140045196
mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.scales BF16 512x640x80 4 32 affine 1192852012 1245280812
mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.weight U32 512x640x320 4 32 affine 194808876 614239276
mtp.safetensors mtp.layers.0.mlp_hyper_connection.block_inject_weight.weight BF16 4x10240 74345100 74427020
mtp.safetensors mtp.layers.0.mlp_hyper_connection.hc_norm.weight BF16 10240 752921132 752941612
mtp.safetensors mtp.layers.0.mlp_hyper_connection.input_mix_weight_down.weight BF16 320x10240 81062540 87616140
1 file name dtype shape quant_bits group_size quant_mode data_start data_end
2492 mtp.safetensors mtp.layers.0.mlp.shared_expert_gate.biases BF16 1x40 4 64 affine 140045196 140045276
2493 mtp.safetensors mtp.layers.0.mlp.shared_expert_gate.scales BF16 1x40 4 64 affine 140045276 140045356
2494 mtp.safetensors mtp.layers.0.mlp.shared_expert_gate.weight U32 1x320 4 64 affine 620915756 620917036
2495 mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.biases BF16 512x2560x20 2 4 64 32 affine 140966956 193395756
2496 mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.scales BF16 512x2560x20 2 4 64 32 affine 641468716 693897516
2497 mtp.safetensors mtp.layers.0.mlp.switch_mlp.down_proj.weight U32 512x2560x80 2 4 64 32 affine 1253145132 1672575532
2498 mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.biases BF16 512x640x80 2 4 64 32 affine 693897772 746326572
2499 mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.scales BF16 512x640x80 2 4 64 32 affine 20277900 72706700
2500 mtp.safetensors mtp.layers.0.mlp.switch_mlp.gate_proj.weight U32 512x640x320 2 4 64 32 affine 760314412 1179744812
2501 mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.biases BF16 512x640x80 2 4 64 32 affine 87616396 140045196
2502 mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.scales BF16 512x640x80 2 4 64 32 affine 1192852012 1245280812
2503 mtp.safetensors mtp.layers.0.mlp.switch_mlp.up_proj.weight U32 512x640x320 2 4 64 32 affine 194808876 614239276
2504 mtp.safetensors mtp.layers.0.mlp_hyper_connection.block_inject_weight.weight BF16 4x10240 74345100 74427020
2505 mtp.safetensors mtp.layers.0.mlp_hyper_connection.hc_norm.weight BF16 10240 752921132 752941612
2506 mtp.safetensors mtp.layers.0.mlp_hyper_connection.input_mix_weight_down.weight BF16 320x10240 81062540 87616140