diff --git a/litellm-rust/loadtest/workloads/sustained-10.json b/litellm-rust/loadtest/workloads/sustained-10.json index b8612a9d554..c25a31a1d2e 100644 --- a/litellm-rust/loadtest/workloads/sustained-10.json +++ b/litellm-rust/loadtest/workloads/sustained-10.json @@ -1,71 +1,20 @@ { "name": "sustained 10 rps for 15 minutes, evenly spaced arrivals, smoke's request mix over every model group", - "stages": [ - { - "duration_s": 900, - "rps": 10 - } - ], + "stages": [{"duration_s": 900, "rps": 10}], "arrivals": "constant", "nonstream_max_s": 360, "warmup_s": 60, "max_in_flight": 30000, - "keys": { - "file": "out/keys.txt", - "zipf_s": 0 - }, - "models": { - "chat": { - "prefix": "gpt-6-astra-", - "count": 10 - }, - "messages": { - "prefix": "claude-opus-5-5-", - "count": 10 - } - }, + "keys": {"file": "out/keys.txt", "zipf_s": 0}, + "models": {"chat": {"prefix": "gpt-6-astra-", "count": 10}, "messages": {"prefix": "claude-opus-5-5-", "count": 10}}, "classes": [ - { - "name": "chat", - "weight": 1, - "endpoint": "chat", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - }, - { - "name": "messages", - "weight": 1, - "endpoint": "messages", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - } + {"name": "chat", "weight": 1, "endpoint": "chat", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}}, + {"name": "messages", "weight": 1, "endpoint": "messages", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}} ] } diff --git a/litellm-rust/loadtest/workloads/sustained-25.json b/litellm-rust/loadtest/workloads/sustained-25.json index 9f29f543bd5..103361af107 100644 --- a/litellm-rust/loadtest/workloads/sustained-25.json +++ b/litellm-rust/loadtest/workloads/sustained-25.json @@ -1,71 +1,20 @@ { "name": "sustained 25 rps for 15 minutes, evenly spaced arrivals, smoke's request mix over every model group", - "stages": [ - { - "duration_s": 900, - "rps": 25 - } - ], + "stages": [{"duration_s": 900, "rps": 25}], "arrivals": "constant", "nonstream_max_s": 360, "warmup_s": 60, "max_in_flight": 30000, - "keys": { - "file": "out/keys.txt", - "zipf_s": 0 - }, - "models": { - "chat": { - "prefix": "gpt-6-astra-", - "count": 10 - }, - "messages": { - "prefix": "claude-opus-5-5-", - "count": 10 - } - }, + "keys": {"file": "out/keys.txt", "zipf_s": 0}, + "models": {"chat": {"prefix": "gpt-6-astra-", "count": 10}, "messages": {"prefix": "claude-opus-5-5-", "count": 10}}, "classes": [ - { - "name": "chat", - "weight": 1, - "endpoint": "chat", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - }, - { - "name": "messages", - "weight": 1, - "endpoint": "messages", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - } + {"name": "chat", "weight": 1, "endpoint": "chat", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}}, + {"name": "messages", "weight": 1, "endpoint": "messages", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}} ] } diff --git a/litellm-rust/loadtest/workloads/sustained-50.json b/litellm-rust/loadtest/workloads/sustained-50.json index de1bd9e1091..c72e0a9332d 100644 --- a/litellm-rust/loadtest/workloads/sustained-50.json +++ b/litellm-rust/loadtest/workloads/sustained-50.json @@ -1,71 +1,20 @@ { "name": "sustained 50 rps for 15 minutes, evenly spaced arrivals, smoke's request mix over every model group", - "stages": [ - { - "duration_s": 900, - "rps": 50 - } - ], + "stages": [{"duration_s": 900, "rps": 50}], "arrivals": "constant", "nonstream_max_s": 360, "warmup_s": 60, "max_in_flight": 30000, - "keys": { - "file": "out/keys.txt", - "zipf_s": 0 - }, - "models": { - "chat": { - "prefix": "gpt-6-astra-", - "count": 10 - }, - "messages": { - "prefix": "claude-opus-5-5-", - "count": 10 - } - }, + "keys": {"file": "out/keys.txt", "zipf_s": 0}, + "models": {"chat": {"prefix": "gpt-6-astra-", "count": 10}, "messages": {"prefix": "claude-opus-5-5-", "count": 10}}, "classes": [ - { - "name": "chat", - "weight": 1, - "endpoint": "chat", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - }, - { - "name": "messages", - "weight": 1, - "endpoint": "messages", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - } + {"name": "chat", "weight": 1, "endpoint": "chat", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}}, + {"name": "messages", "weight": 1, "endpoint": "messages", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}} ] } diff --git a/litellm-rust/loadtest/workloads/sustained-75.json b/litellm-rust/loadtest/workloads/sustained-75.json index 907ba5e2477..74f85df9aed 100644 --- a/litellm-rust/loadtest/workloads/sustained-75.json +++ b/litellm-rust/loadtest/workloads/sustained-75.json @@ -1,71 +1,20 @@ { "name": "sustained 75 rps for 15 minutes, evenly spaced arrivals, smoke's request mix over every model group", - "stages": [ - { - "duration_s": 900, - "rps": 75 - } - ], + "stages": [{"duration_s": 900, "rps": 75}], "arrivals": "constant", "nonstream_max_s": 360, "warmup_s": 60, "max_in_flight": 30000, - "keys": { - "file": "out/keys.txt", - "zipf_s": 0 - }, - "models": { - "chat": { - "prefix": "gpt-6-astra-", - "count": 10 - }, - "messages": { - "prefix": "claude-opus-5-5-", - "count": 10 - } - }, + "keys": {"file": "out/keys.txt", "zipf_s": 0}, + "models": {"chat": {"prefix": "gpt-6-astra-", "count": 10}, "messages": {"prefix": "claude-opus-5-5-", "count": 10}}, "classes": [ - { - "name": "chat", - "weight": 1, - "endpoint": "chat", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - }, - { - "name": "messages", - "weight": 1, - "endpoint": "messages", - "stream_fraction": 0.5, - "context_tokens": { - "dist": "loguniform", - "min": 10000, - "max": 50000 - }, - "ttft_ms": { - "dist": "uniform", - "min": 200, - "max": 800 - }, - "duration_s": { - "dist": "uniform", - "min": 1, - "max": 5 - } - } + {"name": "chat", "weight": 1, "endpoint": "chat", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}}, + {"name": "messages", "weight": 1, "endpoint": "messages", "stream_fraction": 0.5, + "context_tokens": {"dist": "loguniform", "min": 10000, "max": 50000}, + "ttft_ms": {"dist": "uniform", "min": 200, "max": 800}, + "duration_s": {"dist": "uniform", "min": 1, "max": 5}} ] }