date,status,source,server,backend,driver,tool_build,model,quant,api_surface,endpoint_tested,streaming_tested,tool_calling_tested,concurrency,max_context,tokens_per_request,aggregate_tps_mean,mean_ttft_s,p95_ttft_s,repeat_mean_p95_request_mean_decode_interval_ms,setup_friction,best_for,limitations,notes
2026-05-03,measured-local,data/raw/2026-05-03/multi-user,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,not-tested,8,4096,128,161.98,0.307,0.309,47.4,medium,fast local API and multiple tools,manual model management; not a vLLM scheduler,continuous batching; best measured Qwen3.6 balance before plateau
2026-05-03,measured-local,data/raw/2026-05-03/multi-user-coder,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3-Coder 30B-A3B,UD-Q4_K_XL,OpenAI-compatible completions,/v1/completions,yes,not-tested,8,4096,128,173.16,0.382,0.384,43.5,medium,coding assistants and coding agents,manual model management; np16 regressed,continuous batching; best measured Qwen3-Coder balance
2026-05-07,measured-local,data/raw/2026-05-07/latest-stack-rerun/clean-b9049-rerun,Ollama,Vulkan,RADV,0.23.1,Qwen3.6 35B-A3B,Q4_K_M,Ollama API measured; OpenAI-compatible API available,/api/generate,no,not-tested,1,65536,,50.51,,,,easy,beginner chat and Open WebUI,lower short-context speed than direct llama-server,controlled 0.23.1 API warm average across 10 runs; no material change versus 0.21.2
2026-05-03,smoke-test,VLLM_BASELINE.md,kyuz0 vLLM toolbox,ROCm/TheRock,gfx1151,docker.io/kyuz0/vllm-therock-gfx1151:stable; vLLM 0.19.2rc1.dev113,Qwen/Qwen3-0.6B,dtype auto,OpenAI-compatible chat and metrics,/health; /v1/models; /v1/chat/completions; /metrics,no,not-tested,1,2048,,,,,,hard,vLLM server preflight,small model only; no useful throughput claim yet,smoke test proves gfx1151 vLLM serving path works
2026-05-05,smoke-test,data/raw/2026-05-05/server-shootout/llama-server-qwen36-np8-workstation-baseline,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions and chat,/v1/completions; /v1/models; /v1/chat/completions,yes,endpoint-accepted-model-no-call,8,4096,128,173.36,0.252,0.252,44.5,medium,fast local API workflow validation,workstation baseline recorded; VM paused; short 3-rep smoke only,feature probe HTTP 200 for models/completions/chat/stream/tools; 0 throughput errors
2026-05-05,smoke-test,data/raw/2026-05-05/server-shootout/lemonade-b1259-qwen36-np8-workstation-baseline,Lemonade llamacpp-rocm,ROCm,gfx1151,b1259 e77056f; ROCm 7.13.0a20260421,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions and chat,/v1/completions; /v1/models; /v1/chat/completions,yes,endpoint-accepted-model-no-call,8,4096,128,176.43,0.269,0.279,43.7,medium,ROCm llama.cpp workflow validation,workstation baseline recorded; VM paused; short 3-rep smoke only,feature probe HTTP 200 for models/completions/chat/stream/tools; 0 throughput errors; release b1259 published 2026-05-04
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/vulkan-radv/np1,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,endpoint-accepted-model-no-call,1,4096,128,58.80,0.115,0.115,16.2,medium,single-user chat and scripts,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 4096; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/vulkan-radv/np2,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,endpoint-accepted-model-no-call,2,4096,128,96.08,0.156,0.156,19.8,medium,small parallel local API,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 8192; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/vulkan-radv/np4,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,endpoint-accepted-model-no-call,4,4096,128,138.83,0.206,0.207,27.4,medium,small team local API,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 16384; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/vulkan-radv/np8,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions; /v1/models; /v1/chat/completions,yes,endpoint-accepted-model-no-call,8,4096,128,170.87,0.254,0.255,45.2,medium,balanced local API,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 32768; feature probe passed at np16; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/vulkan-radv/np16,llama-server,Vulkan,RADV,b9010 d05fe1d7d,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions and chat,/v1/completions; /v1/models; /v1/chat/completions,yes,endpoint-accepted-model-no-call,16,4096,128,189.72,0.399,0.400,81.9,medium,Vulkan high-concurrency baseline,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 65536; feature probe passed; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/lemonade-b1259-rocm/np1,Lemonade llamacpp-rocm,ROCm,gfx1151,b1259 e77056f; ROCm 7.13.0a20260421,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,endpoint-accepted-model-no-call,1,4096,128,48.62,0.097,0.097,20.0,medium,ROCm single-user comparison,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 4096; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/lemonade-b1259-rocm/np2,Lemonade llamacpp-rocm,ROCm,gfx1151,b1259 e77056f; ROCm 7.13.0a20260421,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,endpoint-accepted-model-no-call,2,4096,128,82.19,0.127,0.132,23.6,medium,ROCm small parallel comparison,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 8192; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/lemonade-b1259-rocm/np4,Lemonade llamacpp-rocm,ROCm,gfx1151,b1259 e77056f; ROCm 7.13.0a20260421,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions,yes,endpoint-accepted-model-no-call,4,4096,128,127.03,0.172,0.175,30.4,medium,ROCm small team comparison,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 16384; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/lemonade-b1259-rocm/np8,Lemonade llamacpp-rocm,ROCm,gfx1151,b1259 e77056f; ROCm 7.13.0a20260421,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions,/v1/completions; /v1/models; /v1/chat/completions,yes,endpoint-accepted-model-no-call,8,4096,128,177.17,0.271,0.284,43.5,medium,high-concurrency local API,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 32768; feature probe passed at np16; 0 throughput errors
2026-05-05,measured-local,data/raw/2026-05-05/server-shootout/full-sweep-qwen36-workstation-baseline/lemonade-b1259-rocm/np16,Lemonade llamacpp-rocm,ROCm,gfx1151,b1259 e77056f; ROCm 7.13.0a20260421,Qwen3.6 35B-A3B,UD-Q4_K_M,OpenAI-compatible completions and chat,/v1/completions; /v1/models; /v1/chat/completions,yes,endpoint-accepted-model-no-call,16,4096,128,207.81,0.460,0.475,74.3,medium,best measured high-concurrency llama.cpp path,workstation baseline; VM paused; 5 measured reps,full sweep row; context_total 65536; feature probe passed; 0 throughput errors
2026-05-05,candidate-not-measured,https://github.com/lemonade-sdk/vllm-rocm/releases/tag/vllm0.20.1-rocm7.12.0-gfx1151,Lemonade vllm-rocm,ROCm,gfx1151,vLLM 0.20.1; ROCm 7.12.0,pending,pending,pending,pending,pending,pending,pending,,,,,,,hard,vLLM serving comparison candidate,not measured on this guide machine yet,release published 2026-05-03
2026-05-07,candidate-not-measured,https://github.com/hec-ovi/vllm-awq4-qwen,hec-ovi vllm-awq4-qwen,ROCm/TheRock,gfx1151,vLLM 0.20.0 patched; TheRock ROCm 7.13 nightly,Qwen3.6-27B,AWQ-INT4 W4A16 g32,OpenAI-compatible chat and responses with tools and vision,/v1/chat/completions; /v1/responses,yes,claimed-by-upstream,1-3,65536-262144,2048,,,,,hard,experimental vLLM AWQ/DFlash server,path not reproduced locally; model downloads and build required; gated drafter path may block,upstream reports DFlash and custom HIP prefill kernel; local guide makes no performance claim yet