{ "date_utc": "2026-05-07", "model": { "hf_id": "unsloth/MiniMax-M2.7-GGUF", "base_hf_id_for_localmaxxing": "unsloth/MiniMax-M2.7", "local_path": "/home/steve/models/minimax-m2.7-ud-iq4_xs-gguf/UD-IQ4_XS/MiniMax-M2.7-UD-IQ4_XS-00001-of-00004.gguf", "quantization": "UD-IQ4_XS", "model_type": "minimax-m2 230B.A10B IQ4_XS - 4.25 bpw", "model_size_bytes": 108405492736, "n_params": 228689764864 }, "hardware": { "gpu": "Intel Arc Pro B70 32GB", "gpu_count": 4, "cpu": "AMD EPYC 9015 8-Core Processor", "os": "Ubuntu 24.04.4 LTS" }, "engine": { "name": "ik_llama.cpp", "commit": "9a26522", "backend": "SYCL RPC over Level Zero", "worker_build": "/home/steve/src/ik_llama.cpp/build-sycl-rpc-b70", "client_build": "/home/steve/src/ik_llama.cpp/build-rpc-client-cpu" }, "first_result": { "label": "ik-rpc-quad-layer-ts0-1111-nofmoe-nommad-syclmultiadd-rtr1-t4-p0n64-nkvo1", "tok_s_out": 13.754201, "tok_s_total": 13.754201, "prompt_tokens": 0, "output_tokens": 64, "context_length": 64, "peak_vram_gb_per_card": 26.4, "jsonl": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/ik-rpc-quad-layer-ts0-1111-nofmoe-nommad-syclmultiadd-rtr1-t4-p0n64-nkvo1-20260507T185647Z.jsonl", "localmaxxing_id": "cmovvoo6f00f5p1017yeb7kxd" }, "best_result": { "label": "ik-rpc-quad-layer-ts0-1111-muge1-fmoe1-fastiq4xsmoe-default-warm-rtr1-t4-ub32-p0n64-nkvo1", "tok_s_out": 14.146009, "tok_s_total": 14.146009, "prompt_tokens": 0, "output_tokens": 64, "context_length": 64, "peak_vram_gb_per_card": 26.4, "jsonl": "/home/steve/bench-results/minimax-m2.7-ud-iq4_xs-gguf/ik-rpc-quad-layer-ts0-1111-muge1-fmoe1-fastiq4xsmoe-default-warm-rtr1-t4-ub32-p0n64-nkvo1-20260507T211315Z.jsonl", "localmaxxing_id": "cmovzx6um00hrp101ldegbaga" }, "best_command": "GGML_DISABLE_FUSED_RMS_NORM=1 GGML_DISABLE_FUSED_MUL_UNARY=1 /home/steve/src/ik_llama.cpp/build-rpc-client-cpu/bin/llama-bench -m /home/steve/models/minimax-m2.7-ud-iq4_xs-gguf/UD-IQ4_XS/MiniMax-M2.7-UD-IQ4_XS-00001-of-00004.gguf -rpc '127.0.0.1:50100|0,127.0.0.1:50101|0,127.0.0.1:50102|0,127.0.0.1:50103|0' -p 0 -n 64 -r 1 -ngl 99 -sm layer -ts 0/1/1/1/1 -fa 0 -nkvo 1 -ub 32 -ctk f16 -ctv f16 -t 4 -w 0 -mmp 1 -rtr 1 -fmoe 1 -no-mmad 0 -muge 1 -v -o json", "sweep": [ { "label": "p0n16-cpukv-norepack", "tok_s_out": 9.977498 }, { "label": "p0n64-gpukv", "tok_s_out": 11.353396 }, { "label": "p0n64-cpukv-norepack-t8", "tok_s_out": 12.564798 }, { "label": "p0n64-rtr1-t8", "tok_s_out": 13.465415 }, { "label": "p0n64-rtr1-t16", "tok_s_out": 12.826655 }, { "label": "p0n64-rtr1-t4", "tok_s_out": 13.754201 }, { "label": "p0n64-rtr1-t2", "tok_s_out": 13.671846 }, { "label": "p0n64-rtr1-t4-ub64", "tok_s_out": 13.690724 }, { "label": "p0n64-rtr1-t4-ub16", "tok_s_out": 13.65025 }, { "label": "p0n64-fused-mulmultiadd-rtr1-t4-ub32", "tok_s_out": 12.330226, "note": "Experimental fused SYCL MUL_MULTI_ADD regressed versus decomposed mul + MULTI_ADD." }, { "label": "p0n1-fmoe1-syclmoe-nommad-rtr1-t4-ub32", "tok_s_out": 1.440125, "note": "Smoke test after adding SYCL MOE_FUSED_UP_GATE support." }, { "label": "p0n64-fmoe1-syclmoe-nommad-rtr1-t4-ub32", "tok_s_out": 13.839857 }, { "label": "p0n64-fmoe1-syclmoe-syclmulmultiadd-rtr1-t4-ub32", "tok_s_out": 13.899895 }, { "label": "p0n64-muge1-fmoe1-syclmoe-syclmulmultiadd-rtr1-t4-ub32", "tok_s_out": 13.989985, "note": "Conservative fused-MoE best; accepted by LocalMaxxing as cmovxb67400g6p10100by6frn." }, { "label": "p0n1-muge1-fmoe1-syclmoe-timing-warm-rtr1-t4-ub32", "tok_s_out": 5.101027, "note": "Warm timing probe. MOE_FUSED_UP_GATE total was about 34 ms across 62 calls; MUL_MAT_ID down path about 10 ms across 62 calls." }, { "label": "graph-fit-ts0-1111-muge1-fmoe1-rtr1-t4-ub32-p0n1", "tok_s_out": null, "note": "Failed: split vector included CPU as a nonzero target, so auto-fit tried to offload output to CPU device 4 with 0 MiB free." }, { "label": "graph-fit-ts11110-muge1-fmoe1-rtr1-t4-b32ub32-p0n1", "tok_s_out": null, "note": "Failed: corrected split reached graph buffer planning but one worker attempted a 33904507904 byte SYCL allocation on one B70; removing -muge did not fix it." }, { "label": "p0n64-muge1-fmoe1-fastiq4xsmoe-rtr1-t4-ub32", "tok_s_out": 14.130936, "note": "First full run with direct IQ4_XS merged up/gate active-expert fast path." }, { "label": "p0n64-muge1-fmoe1-fastiq4xsmoe-fastdown-rtr1-t4-ub32", "tok_s_out": 13.207767, "note": "Negative result: experimental IQ4_XS MUL_MAT_ID down-projection fast path regressed, so it is disabled by default behind GGML_SYCL_FAST_MUL_MAT_ID_IQ4_XS=1." }, { "label": "p0n64-muge1-fmoe1-fastiq4xsmoe-default-warm-rtr1-t4-ub32", "tok_s_out": 14.146009, "note": "Current best; warm-worker run with direct IQ4_XS merged up/gate active-expert fast path enabled and down fast path disabled. Accepted by LocalMaxxing as cmovzx6um00hrp101ldegbaga." } ], "patches": [ "patches/ik-llama-minimax-rpc-sycl-20260507.patch" ], "blocked_or_negative": [ { "flag": "-fa 1", "failure": "FLASH_ATTN_EXT unsupported in SYCL RPC worker" }, { "flag": "-sm graph", "failure": "incorrect ts0-1111 split tried to use CPU as a split device; corrected ts11110 split attempted one 33904507904 byte SYCL allocation on one B70" }, { "flag": "-fmoe 1 -no-mmad 1", "failure": "fixed by local SYCL MOE_FUSED_UP_GATE support; old failure retained for history" }, { "flag": "-t 2,4,8 with -muge 1 -rtr 1", "failure": "stopped making progress during llama_repack_up_gate_exps around layer 56; single -t 4 completed" } ], "next_steps": [ "Investigate graph split buffer planning and RPC allocation behavior so MiniMax can use more than layer-split memory sharding.", "Profile the direct IQ4_XS up/gate path with GGML_SYCL_OP_TIMING=1 and keep only if it stays warm-run positive.", "Rework the IQ4_XS down-projection fast path before enabling it again.", "Study external CUDA/ROCm MiniMax multi-GPU deployments for graph/tensor split ideas that can be ported to SYCL RPC." ] }