{ "schema": "anvil-serving.controlled-ab-configuration/v1", "hardware": "2x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96 GB each", "measurement_path": "Direct loopback API on the same serving host; primary router tier quiesced", "source_revision": "c721ad95cec8d9bd73b4ae762ca568fbf976c9e1", "profiles": { "baseline": { "repository": "brandonmusic/GLM-5.3-Flash-tr3-4bpw", "revision": "a5fee929cf4888b1824323e33e8a19b60129e025", "recipe_digest": "54172cd87da0e28180a4aa2dd4a0c10355d45fae328fbfc622c0345286c7b65b", "registry_digest": "3beba851889013e16084ccb5d6154d9dd77c4b742a25620a11bd68686f6afab6", "image": "verdictai/glm53-flash-exl3-k4:r19-sm120-tp2-ep2-dcp2-v84-language-only@sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692", "served_model_name": "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "context_tokens": 327680, "memory_limit_mib": 53248, "memory_swap_limit_mib": 53249, "host_memory_reserve_mib": 12288, "flags": [ "--served-model-name glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec", "--host 0.0.0.0", "--port 30000", "--chat-template /opt/glm53/chat_template.multimodal.jinja", "--limit-mm-per-prompt '{\"image\":8,\"video\":0}'", "--tensor-parallel-size 2", "--enable-expert-parallel", "--decode-context-parallel-size 1", "--dcp-comm-backend a2a", "--dtype bfloat16", "--load-format safetensors", "--moe-backend b12x", "--attention-backend FLASHINFER_MLA_SPARSE_SM120", "--kv-cache-dtype fp8_ds_mla", "--max-model-len 327680", "--max-num-batched-tokens 2048", "--max-num-seqs 4", "--gpu-memory-utilization 0.970", "--enable-chunked-prefill", "--enable-prefix-caching", "--generation-config /root/.cache/huggingface", "--reasoning-parser glm45", "--reasoning-config '{}'", "--tool-call-parser glm47", "--enable-auto-tool-choice", "--disable-custom-all-reduce" ], "quantization": "uniform K4 EXL3/TR3 routed experts; native-precision non-routed tensors; FP8 DS-MLA KV; speculation disabled; vision enabled" }, "candidate": { "repository": "nvidia/GLM-5.3-Flash-FP4-Local", "revision": "76348ddb2be2e35e39ecfc00e75dc07703f4f3d5", "recipe_digest": "e2b8def91989109f8cfa9e95a9f8102f1ab954fafcc9b8f42bacc330690130eb", "registry_digest": "773394b53c51728df11f1369ce32eed12b5a10985cefba3fdbc8815063d153b3", "image": "ghcr.io/local-inference-lab/vllm:release-10152026@sha256:745feb916ee816f5409fd58386114be4f22fbede7f9c0e0a7b8c4fe8dac0fb29", "served_model_name": "glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "context_tokens": 327680, "memory_limit_mib": 53248, "memory_swap_limit_mib": 53248, "host_memory_reserve_mib": 12288, "flags": [ "--host 0.0.0.0", "--port 30000", "--additional-config '{\"glm53_kda_decode_backend\":\"auto\",\"kda_prefill_backend\":\"b12x\"}'", "--attention-backend B12X", "--block-size 256", "--chat-template /opt/lil/runtime/templates/glm53-flash.jinja", "--compilation-config '{\"cudagraph_mode\":\"FULL_AND_PIECEWISE\"}'", "--cp-kv-cache-interleave-size 4", "--cudagraph-capture-sizes 1 2 4 8 12 16 20 24 28 32", "--dcp-kv-cache-interleave-size 4", "--decode-context-parallel-size 2", "--default-chat-template-kwargs '{\"reasoning_effort\":\"max\",\"clear_thinking\":false}'", "--dtype bfloat16", "--enable-auto-tool-choice", "--enable-chunked-prefill", "--no-enable-flashinfer-autotune", "--enable-prefix-caching", "--enable-prompt-tokens-details", "--gpu-memory-utilization 0.985", "--kv-cache-dtype fp8", "--kv-cache-memory-bytes 7650410496", "--linear-backend b12x", "--load-format instanttensor", "--mamba-cache-mode align", "--max-cudagraph-capture-size 32", "--max-model-len 327680", "--max-num-batched-tokens 4096", "--max-num-seqs 4", "--max-parallel-prefills 1", "--moe-backend b12x", "--override-generation-config '{\"temperature\":1.0,\"top_p\":0.95}'", "--prefill-compute-share 0.4", "--prefill-schedule-interval 1", "--quantization modelopt_mixed", "--reasoning-parser glm45", "--recurrent-checkpoint-policy request_boundaries", "--speculative-config '{\"method\":\"mtp\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"standard\",\"moe_backend\":\"b12x\",\"attention_backend\":\"B12X\",\"num_speculative_tokens\":3}'", "--tensor-parallel-size 2", "--tool-call-parser glm47", "--served-model-name glm53-nvidia-fp4-local-tp2-327k-c4-mtp3-r2", "--revision 76348ddb2be2e35e39ecfc00e75dc07703f4f3d5" ], "quantization": "NVIDIA QAD NVFP4/MXFP8 compressed scales; FP8 KV; MTP3", "environment": [ "B12X_DYNAMIC_DIRECT_EXPERT_SCALES=1", "B12X_DYNAMIC_SKIP_SPLIT_BARRIER_RESET=1", "B12X_DYNAMIC_SPLIT_COMPUTE_MAC=224", "B12X_DYNAMIC_SPLIT_FAST_PREPARE=1", "B12X_DYNAMIC_SPLIT_LOW_SMEM=1", "B12X_DYNAMIC_SPLIT_ROUTE_COMPUTE=1", "B12X_DYNAMIC_WORK_SOURCE=persistent_grid", "B12X_MHC_PDL=1", "B12X_PCIE_ONESHOT_BLOCK_LIMIT=4", "B12X_PCIE_ONESHOT_PDL=1", "B12X_PCIE_ONESHOT_THREADS=512", "B12X_W4A16_A4_PREFILL=1", "B12X_W4A16_FP32_TOPK_WEIGHTS=1", "CUBLAS_WORKSPACE_CONFIG=:4096:1", "CUDA_DEVICE_ORDER=PCI_BUS_ID", "CUTE_DSL_ARCH=sm_120a", "GLOO_SOCKET_IFNAME=lo", "INSTANTTENSOR_BACKEND=BUFFERED", "NCCL_BUFFSIZE=1048576", "NCCL_IB_DISABLE=1", "NCCL_MAX_NCHANNELS=2", "NCCL_MIN_NCHANNELS=2", "NCCL_NET_PLUGIN=none", "NCCL_P2P_LEVEL=SYS", "NCCL_PROTO=LL,LL128,Simple", "NCCL_SOCKET_IFNAME=lo", "NCCL_TUNER_PLUGIN=none", "OMP_NUM_THREADS=1", "PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,large_segment_size_mb:12", "SAFETENSORS_FAST_GPU=1", "TORCH_CUDA_ARCH_LIST=12.0a", "VLLM_B12X_MLA_CKV_GATHER=1", "VLLM_B12X_MLA_CKV_GATHER_MAX_TOKENS=65536", "VLLM_B12X_MOE_FP4_FORCE_A16=1", "VLLM_CAUSAL_CONV1D_UPDATE_HOIST=1", "VLLM_DISABLED_KERNELS=MarlinFP8ScaledMMLinearKernel", "VLLM_DISABLE_SHARED_EXPERTS_STREAM=0", "VLLM_ENABLE_PCIE_ALLREDUCE=1", "VLLM_GLM53_DFLASH_ATTN=1", "VLLM_GLM53_EMBED_HOST=1", "VLLM_GLM53_KDA_GATE_SIDE_STREAM=1", "VLLM_GLM53_KDA_MXFP8_ACTIVATION_MODE=prefill", "VLLM_GLM53_L2_PREFETCH=1", "VLLM_GLM53_L2_PREFETCH_PERSIST_MB=0", "VLLM_GLM53_MTP_DRAFT_HEAD=bf16", "VLLM_GLM53_ONLINE_DENSE_MXFP8=0", "VLLM_GLM53_SPLIT_MAMBA_BLOCK_SIZE=auto", "VLLM_GLM53_SPLIT_TARGET_BLOCK_SIZE=2048", "VLLM_LM_HEAD_A16=1", "VLLM_MTP_NVFP4_LM_HEAD=0", "VLLM_MXFP8_LM_HEAD=0", "VLLM_PCIE_ALLREDUCE_BACKEND=b12x", "VLLM_PCIE_DMA_MIN_BYTES=6MB", "VLLM_PCIE_TWOSHOT_ALLREDUCE_MAX_SIZE=off", "VLLM_PLUGINS=", "VLLM_SHARE_PYNCCL_COMMS=1", "VLLM_USE_FASTOKENS=1", "VLLM_USE_FLASHINFER_SAMPLER=1", "VLLM_USE_V2_MODEL_RUNNER=1", "VLLM_WORKER_MULTIPROC_METHOD=spawn", "HF_HUB_OFFLINE=1", "TRANSFORMERS_OFFLINE=1" ], "residency_note": "Host embeddings enabled by the upstream recipe; this is not a strict GPU-only residency claim." } }, "limits": [ "Different quantization, runtime, attention backend, DCP, batching and speculative settings: compare complete configurations, not an isolated MTP or quantization effect.", "Candidate .985 memory-utilization flag is retained but explicit KV bytes bypass its profiling." ], "redactions": "Private owner identity, topology, container IDs, GPU UUIDs and operator file paths omitted. Container-internal public model/runtime paths are retained. Exact private registry digests identify the original configurations, not a hash of this projection. Actual per-profile endpoint ports are projected to synthetic30000; this projection is not the original launch byte sequence." }