[Rust][Benchmark] Support HF ShareGPT datasets in multi-turn mode

rust/src/bench/README.md

@@ -664,6 +664,7 @@ With `--multi-turn`, `--num-prompts` controls the number of **conversations**, n
 
 - `--dataset-name random` — synthetic conversations with controllable per-turn token lengths. Auto-sets `min_tokens` to enforce output length without `ignore_eos`.
 - `--dataset-name sharegpt` — loads all turns (not just the first two); filters for entries with ≥ 2 real turns.
+- `--dataset-name hf` — downloads a ShareGPT-format HuggingFace config, then loads it like `sharegpt`.
 
 **Prefix sharing** (random dataset): when `--multi-turn-prefix-global-ratio` or `--multi-turn-prefix-conversation-ratio` is > 0, each turn sends a fixed-length message (no history accumulation) composed of a global prefix + per-conversation prefix + unique suffix. The two ratios must sum to < 1.0.
 

rust/src/bench/src/config.rs

@@ -548,13 +548,15 @@ impl BenchConfig {
                 ));
             }
 
-            // Normalize and validate min/max turns. ShareGPT only consumes max_turns
-            // (the loader walks all available turns up to the cap), so the
-            // min/num/max coupling used for synthetic generation does not apply.
-            if args.dataset_name == DatasetName::ShareGpt {
+            // Normalize and validate min/max turns. Conversation datasets only
+            // consume max_turns (the loader walks all available turns up to the
+            // cap), so the min/num/max coupling used for synthetic generation
+            // does not apply.
+            if matches!(args.dataset_name, DatasetName::ShareGpt | DatasetName::Hf) {
                 if args.multi_turn_max_turns == 1 {
                     return Err(BenchError::Config(
-                        "--multi-turn-max-turns must be at least 2 for ShareGPT multi-turn".into(),
+                        "--multi-turn-max-turns must be at least 2 for ShareGPT-format multi-turn"
+                            .into(),
                     ));
                 }
             } else {
@@ -712,7 +714,7 @@ impl BenchConfig {
             multi_turn_min_turns,
             multi_turn_max_turns,
             sharegpt_multi_turn_max_turns: if args.multi_turn
-                && args.dataset_name == DatasetName::ShareGpt
+                && matches!(args.dataset_name, DatasetName::ShareGpt | DatasetName::Hf)
                 && args.multi_turn_max_turns != 0
             {
                 Some(args.multi_turn_max_turns)
@@ -1040,6 +1042,30 @@ mod tests {
         assert_eq!(config.sharegpt_multi_turn_max_turns, Some(20));
     }
 
+    #[test]
+    fn test_hf_multi_turn_uses_conversation_turn_cap() {
+        let args = vec![
+            "vllm-bench",
+            "--backend",
+            "openai-chat",
+            "--multi-turn",
+            "--model",
+            "test-model",
+            "--dataset-name",
+            "hf",
+            "--dataset-path",
+            "org/sharegpt-dataset",
+            "--hf-subset",
+            "sharegpt",
+            "--multi-turn-max-turns",
+            "2",
+        ];
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
+
+        assert_eq!(config.sharegpt_multi_turn_max_turns, Some(2));
+    }
+
     #[test]
     fn test_sweep_summary_percentiles_default_empty() {
         let args = base_multi_turn_args();

rust/src/bench/src/multi_turn.rs

@@ -175,9 +175,29 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
             ));
         }
         DatasetName::Hf => {
-            return Err(BenchError::Config(
-                "HF dataset multi-turn is not yet supported. Use 'random' or 'sharegpt' with --multi-turn.".into(),
-            ));
+            let tok = tokenizer
+                .as_ref()
+                .ok_or_else(|| BenchError::Config("HF dataset requires a tokenizer".into()))?;
+            let dataset_id = config.dataset_path.as_deref().ok_or_else(|| {
+                BenchError::Config("--dataset-path is required for --dataset-name hf".into())
+            })?;
+            let (downloaded_path, _config, _split) =
+                crate::datasets::hf_dataset::download_hf_dataset(
+                    dataset_id,
+                    config.hf_subset.as_deref(),
+                    config.hf_split.as_deref(),
+                    config.num_prompts,
+                )
+                .await?;
+            crate::datasets::multi_turn::load_sharegpt_multi_turn(
+                tok,
+                &downloaded_path,
+                config.num_prompts,
+                config.hf_output_len,
+                config.sharegpt_multi_turn_max_turns,
+                config.seed,
+                &config.request_id_prefix,
+            )?
         }
     };