data: add 41 measured benchmark rows (fp8/nvfp4/bf16 x TP1/TP2 x offload/hybrid)
Browse files- benchmarks.jsonl +41 -0
benchmarks.jsonl
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 1, "moe_backend": "offload", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 93.9, "value_max": null, "unit": "tok/s", "notes": "warm; 399 tokens; single RTX 3090"}
|
| 2 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 1, "moe_backend": "offload", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "ttft_s", "target": "", "value": 4.03, "value_max": null, "unit": "s", "notes": "warm"}
|
| 3 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 1, "moe_backend": "offload", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 90.5, "value_max": null, "unit": "tok/s", "notes": "cold; includes FlashInfer JIT"}
|
| 4 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 1, "moe_backend": "offload", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "ttft_s", "target": "", "value": 79.4, "value_max": null, "unit": "s", "notes": "cold; JIT compile dominated"}
|
| 5 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 1, "moe_backend": "hybrid", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 61.0, "value_max": null, "unit": "tok/s", "notes": "warm; 399 tokens"}
|
| 6 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 1, "moe_backend": "hybrid", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "ttft_s", "target": "", "value": 2.28, "value_max": null, "unit": "s", "notes": "warm"}
|
| 7 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 1, "moe_backend": "hybrid", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 2000, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 56.4, "value_max": null, "unit": "tok/s", "notes": "cold"}
|
| 8 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 1, "moe_backend": "offload", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 3600, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 86.0, "value_max": 106.0, "unit": "tok/s", "notes": "range across batches; flat vs 2000 MT/s \u2014 PCIe-bound, no fp8 CPU path"}
|
| 9 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 1, "moe_backend": "hybrid", "attention_backend": "hybrid_linear", "extra_flags": "", "ram_mts": 3600, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 76.0, "value_max": 80.9, "unit": "tok/s", "notes": "+29-31% vs 2000 MT/s \u2014 CPU executor benefits from RAM bandwidth"}
|
| 10 |
+
{"date": "2026-08-27", "engine": "PR#104 3c8b281 + f1ff8e2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "attention_backend": "triton", "ram_mts": 3600, "category": "serving", "moe_backend": "hybrid", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "metric": "decode_tok_s", "target": "", "value": 2.2, "value_max": null, "unit": "tok/s", "notes": "pre-affinity-fix; both CPU executors pinned to same cores 0..16"}
|
| 11 |
+
{"date": "2026-08-27", "engine": "PR#104 3c8b281 + f1ff8e2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "attention_backend": "triton", "ram_mts": 3600, "category": "serving", "moe_backend": "hybrid", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7 --moe-cpu-threads 8", "metric": "decode_tok_s", "target": "", "value": 8.62, "value_max": 9.47, "unit": "tok/s", "notes": "workaround: cap threads per rank; 4.3x over broken"}
|
| 12 |
+
{"date": "2026-08-27", "engine": "PR#104 3c8b281 + f1ff8e2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "attention_backend": "triton", "ram_mts": 3600, "category": "serving", "moe_backend": "offload", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "metric": "decode_tok_s", "target": "", "value": 3.1, "value_max": null, "unit": "tok/s", "notes": "clean rerun; 219 tokens in 70.3s; earlier contaminated measurement (stale server on GPUs) discarded"}
|
| 13 |
+
{"date": "2026-08-27", "engine": "PR#104 3c8b281 + f1ff8e2", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 2, "moe_backend": "offload", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "load_test", "metric": "load_status", "target": "", "value": "oom_at_load", "value_max": null, "unit": "", "notes": "62 GB RAM insufficient for ~35 GB pinned expert banks + runtime"}
|
| 14 |
+
{"date": "2026-08-26", "engine": "PR#104 3c8b281", "model": "Qwen3.6-35B-A3B-BF16", "quant": "bf16", "tp_size": 2, "moe_backend": "offload", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "load_test", "metric": "load_status", "target": "", "value": "oom_at_load", "value_max": null, "unit": "", "notes": "67 GB model vs 62 GB RAM; crashed at expert load (60/62 GB used)"}
|
| 15 |
+
{"date": "2026-08-28", "engine": "PR#104 + f1ff8e2 (lm_head fix)", "model": "nvidia/Qwen3.6-35B-A3B-FP8", "quant": "fp8", "tp_size": 2, "moe_backend": "hybrid", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "load_test", "metric": "load_status", "target": "", "value": "pass", "value_max": null, "unit": "", "notes": "lm_head vocab-sharding fix verified; passes previous crash point"}
|
| 16 |
+
{"date": "2026-08-28", "engine": "PR#104 + f1ff8e2 (lm_head fix)", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "moe_backend": "hybrid", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "load_test", "metric": "load_status", "target": "", "value": "pass", "value_max": null, "unit": "", "notes": "lm_head vocab-sharding fix verified; passes previous crash point"}
|
| 17 |
+
{"date": "2026-08-31", "engine": "PR#104 + c23ec78 (core-partition fix)", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "moe_backend": "hybrid", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 32.7, "value_max": null, "unit": "tok/s", "notes": "client wall; 219 tokens in 6.7s; disjoint pins rank0 0..14 / rank1 1..15"}
|
| 18 |
+
{"date": "2026-08-31", "engine": "PR#104 + c23ec78 (core-partition fix)", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "moe_backend": "hybrid", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "serving", "metric": "decode_tok_s", "target": "", "value": 59.26, "value_max": null, "unit": "tok/s", "notes": "engine-instantaneous per-batch gen throughput"}
|
| 19 |
+
{"date": "2026-08-27", "engine": "PR#104 + f1ff8e2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 1, "moe_backend": "hybrid", "attention_backend": "triton", "extra_flags": "", "ram_mts": 3600, "category": "capacity", "metric": "kv_usable_tokens", "target": "", "value": 262144, "value_max": null, "unit": "tokens", "notes": "baseline"}
|
| 20 |
+
{"date": "2026-08-27", "engine": "PR#104 + f1ff8e2", "model": "nvidia/Qwen3.6-35B-A3B-NVFP4", "quant": "nvfp4", "tp_size": 2, "moe_backend": "hybrid", "attention_backend": "triton", "extra_flags": "--tensor-parallel-size 2 --attention-backend triton --expert-load serial --disable-pynccl --memory-ratio 0.7", "ram_mts": 3600, "category": "capacity", "metric": "kv_usable_tokens", "target": "", "value": 524288, "value_max": null, "unit": "tokens", "notes": "2.4x KV claim confirmed at TP=2"}
|
| 21 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_stream_read", "value": 54.1, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 22 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 3600, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_stream_read", "value": 80.8, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 23 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_h2d", "value": 8.6, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 24 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 3600, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_h2d", "value": 8.5, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 25 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_nvfp4", "value": 28.1, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 26 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 3600, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_nvfp4", "value": 46.7, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 27 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_bf16", "value": 41.5, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 28 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 3600, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_bf16", "value": 71.3, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 29 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_ds_fp4", "value": 33.4, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 30 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 3600, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_ds_fp4", "value": 49.2, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 31 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_mxfp4", "value": 21.2, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 32 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 3600, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_mxfp4", "value": 35.2, "value_max": null, "unit": "GB/s", "notes": ""}
|
| 33 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "bf16", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_effective", "value": 50.5, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 34 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "bf16", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_gather", "value": 6.6, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 35 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "nvfp4", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_effective", "value": 48.6, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 36 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "nvfp4", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_gather", "value": 6.9, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 37 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "fp8", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_gather", "value": 8.0, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: offload"}
|
| 38 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "mxfp4", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_effective", "value": 28.8, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 39 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "mxfp4", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_gather", "value": 6.0, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 40 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "ds_fp4", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "cpu_moe_effective", "value": 47.7, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|
| 41 |
+
{"date": "2026-08-22", "engine": "v0.1.2", "model": "", "quant": "ds_fp4", "tp_size": 0, "moe_backend": "", "attention_backend": "", "extra_flags": "ft bench bw", "ram_mts": 2000, "category": "bandwidth", "metric": "bandwidth_gbs", "target": "pcie_gather", "value": 6.4, "value_max": null, "unit": "GB/s", "notes": "auto-picked backend: hybrid"}
|