[FEAT] add rampup for GuideLLM concurrent benchmarks

README.md

@@ -38,6 +38,7 @@ The fork baseline is commit [`3c4264d`](https://github.com/InseeFrLab/auto-tunin
 ### Benchmarking (GuideLLM)
 
 - **`benchmark.warmup` / `benchmark.cooldown`** — exclude cold-start and shutdown phases from reported metrics to reduce variance ([#24](https://github.com/InseeFrLab/auto-tuning-vllm/pull/24))
+- **`benchmark.rampup`** — linear load ramp-up duration in seconds before reaching target concurrency
 - **`benchmark.sample_requests`** — control per-request samples in benchmark JSON output (default `0` keeps files small; requires GuideLLM `>= 0.5.4`) ([#27](https://github.com/InseeFrLab/auto-tuning-vllm/pull/27))
 
 ### Bug fixes

auto_tune_vllm/benchmarks/config.py

@@ -34,6 +34,9 @@ class BenchmarkConfig:
     warmup: Optional[float] = None
     cooldown: Optional[float] = None
 
+    # GuideLLM ramp-up duration in seconds (linear increase to target rate). Omit to disable.
+    rampup: Optional[float] = None
+
     # Max detailed request samples stored in benchmark JSON output (GuideLLM --sample-requests).
     sample_requests: int = 0
 
@@ -42,7 +45,11 @@ class BenchmarkConfig:
     logging_level: Literal["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"] = "INFO"
 
     def __post_init__(self) -> None:
-        for name, value in (("warmup", self.warmup), ("cooldown", self.cooldown)):
+        for name, value in (
+            ("warmup", self.warmup),
+            ("cooldown", self.cooldown),
+            ("rampup", self.rampup),
+        ):
             if value is None:
                 continue
             if value <= 0:

auto_tune_vllm/benchmarks/providers.py

@@ -338,6 +338,8 @@ def _build_guidellm_command(
             cmd.extend(["--warmup", str(config.warmup)])
         if config.cooldown is not None:
             cmd.extend(["--cooldown", str(config.cooldown)])
+        if config.rampup is not None:
+            cmd.extend(["--rampup", str(config.rampup)])
 
         # Add dataset or synthetic data configuration
         if config.use_synthetic_data:

docs/configuration.md

@@ -306,6 +306,14 @@ When both `warmup` and `cooldown` are fractional, their sum must stay below `1`
 #### `cooldown` (number, optional)
 GuideLLM cooldown period at the end of the run, also excluded from metrics. Same format as `warmup` (e.g. `0.1` for the last 10%).
 
+#### `rampup` (number, optional)
+GuideLLM ramp-up duration in seconds. Requests are spread linearly from zero up to the target concurrency (`rate`) over this period at the start of the benchmark.
+
+- **Seconds**: e.g. `rampup: 10` = 10 seconds to reach full concurrency.
+- Omit or set to `null` to disable (GuideLLM default).
+
+Unlike `warmup`, ramp-up requests are included in reported metrics — it controls how load increases, not which phase is measured. See [GuideLLM benchmark docs](https://github.com/vllm-project/guidellm/blob/main/docs/getting-started/benchmark.md).
+
 #### `sample_requests` (integer, optional)
 Maximum number of detailed request samples stored in GuideLLM benchmark JSON output (`--sample-requests`). Default: `0` (no per-request samples; keeps result files small). Set to a positive value when you need request-level timings for debugging or deeper analysis. Requires [GuideLLM](https://github.com/vllm-project/guidellm) `>= 0.5.4` (fix for `--sample-requests` in [v0.5.4](https://github.com/vllm-project/guidellm/releases/tag/v0.5.4)).
 
@@ -318,6 +326,7 @@ benchmark:
   rate: 16
   warmup: 0.1
   cooldown: 0.1
+  rampup: 10
   # sample_requests: 20  # optional; default 0
 ```
 
@@ -691,6 +700,7 @@ benchmark:
   rate: 100
   # warmup: 0.1
   # cooldown: 0.1
+  # rampup: 10
   # sample_requests: 0  # default; increase to keep detailed request samples in benchmark JSON
 
 logging:

examples/study_config.yaml

@@ -31,6 +31,7 @@ benchmark:
   # Optional: reduce metric variance (requires recent GuideLLM CLI)
   # warmup: 0.1
   # cooldown: 0.1
+  # rampup: 10
   # sample_requests: 0  # default; increase to keep detailed request samples in benchmark JSON
 
 logging:

examples/study_config_minimal.yaml

@@ -31,6 +31,7 @@ benchmark:
   output_tokens: 100
   # warmup: 0.1
   # cooldown: 0.1
+  # rampup: 10
   # sample_requests: 0  # default; increase to keep detailed request samples in benchmark JSON
 
 # No logging section specified - will use console logging only

tests/benchmarks/test_guidellm_command.py

@@ -50,3 +50,23 @@ def test_fraction_warmup_plus_cooldown_must_leave_measured_window():
 
 def test_mixed_fraction_and_absolute_skips_sum_check():
     BenchmarkConfig(model="m", warmup=0.1, cooldown=10)
+
+
+def test_rampup_included_when_set():
+    cmd = _build_cmd(rampup=10)
+    assert cmd[cmd.index("--rampup") + 1] == "10"
+
+
+def test_rampup_omitted_when_unset():
+    cmd = _build_cmd()
+    assert "--rampup" not in cmd
+
+
+def test_rampup_zero_rejected():
+    with pytest.raises(ValueError, match="rampup"):
+        BenchmarkConfig(model="m", rampup=0)
+
+
+def test_negative_rampup_rejected():
+    with pytest.raises(ValueError, match="rampup"):
+        BenchmarkConfig(model="m", rampup=-1)