diff --git a/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env b/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env
index 7e02106..d69c72d 100644
--- a/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env
+++ b/experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env
@@ -90,18 +90,20 @@ API_SERVER_COUNT="${API_SERVER_COUNT:-1}"
 
 # Per-TP parameter overrides (applied in start_vllm_docker.sh via case $TP).
 # DSV4-Flash w8a8 weight per die = ~280GiB / TP. Each die has 64GB HBM.
+# Defaults below are the verified-OK values (from the 2026-07-29..30 adaptive
+# runs' server_cmd.txt); 1048576-context OOMs on TP4/TP8 (KV cache budget).
 #   TP=4:  ~39 GiB/die weights -> 25 GiB for KV cache; cap context to avoid OOM
 #   TP=8:  ~35 GiB/die weights -> 29 GiB for KV cache; balanced
 #   TP=16: ~17.5 GiB/die weights -> 46 GiB for KV cache; full context
 TP4_GPU_MEMORY_UTILIZATION="${TP4_GPU_MEMORY_UTILIZATION:-0.9}"
-TP4_MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-1048576}"
-TP4_MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-64}"
+TP4_MAX_MODEL_LEN="${TP4_MAX_MODEL_LEN:-32768}"
+TP4_MAX_NUM_SEQS="${TP4_MAX_NUM_SEQS:-128}"
 TP8_GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.9}"
-TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-1048576}"
-TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-64}"
+TP8_MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-65536}"
+TP8_MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-256}"
 TP16_GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.9}"
-TP16_MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-1048576}"
-TP16_MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-64}"
+TP16_MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-131072}"
+TP16_MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-256}"
 
 # vLLM-Ascend-specific launch flags injected by start_vllm_docker.sh.
 # attention backend for 910C: use the fused/atb attention path. Adjust per image.
diff --git a/src/sskj/bench/cli.py b/src/sskj/bench/cli.py
index f5d7db5..a93ea27 100644
--- a/src/sskj/bench/cli.py
+++ b/src/sskj/bench/cli.py
@@ -87,13 +87,16 @@ def _cmd_run(args: argparse.Namespace, root: Path) -> int:
     url = args.url or f"http://127.0.0.1:{default_port}"
     base_url, host, port = _normalize_url(url)
 
+    # API model name: SERVED_MODEL_NAME first (vLLM validates it strictly);
+    # tokenizer for prompt length counting stays on the local MODEL_PATH.
     model = (
         args.model
-        or experiment_env.get("MODEL_PATH")
         or experiment_env.get("SERVED_MODEL_NAME")
+        or experiment_env.get("MODEL_PATH")
         or experiment_env.get("MODEL_NAME")
         or "model"
     )
+    tokenizer = args.model or experiment_env.get("MODEL_PATH")
     dataset = args.dataset or experiment_env.get("DATASET_PATH")
     client_mode = bench_config.resolve_client_mode(args.client, args.platform, experiment_env)
 
@@ -189,6 +192,7 @@ def _cmd_run(args: argparse.Namespace, root: Path) -> int:
             host=host,
             port=port,
             model=model,
+            tokenizer=tokenizer,
             dataset_name=dataset_name,
             dataset_path=dataset,
             container_dataset_path=experiment_env.get("CONTAINER_DATASET_PATH"),
diff --git a/src/sskj/bench/runner.py b/src/sskj/bench/runner.py
index 3c2732a..67cc517 100644
--- a/src/sskj/bench/runner.py
+++ b/src/sskj/bench/runner.py
@@ -18,6 +18,7 @@ class BenchClientOptions:
     host: str
     port: int
     model: str
+    tokenizer: str | None = None
     dataset_name: str = "random"
     dataset_path: str | None = None
     container_dataset_path: str | None = None
@@ -45,6 +46,10 @@ def _bench_args(scenario: dict[str, Any], options: BenchClientOptions) -> list[s
         str(options.port),
         "--model",
         options.model,
+    ]
+    if options.tokenizer:
+        args += ["--tokenizer", options.tokenizer]
+    args += [
         "--dataset-name",
         options.dataset_name,
         "--random-input-len",
@@ -107,8 +112,9 @@ def run_scenario(scenario: dict[str, Any], options: BenchClientOptions) -> int:
         cmd = ["docker", "run", "--rm", "--network", "host"]
         if options.root:
             cmd += ["-v", f"{options.root}:{options.root}"]
-        if Path(options.model).exists():
-            cmd += ["-v", f"{options.model}:{options.model}:ro"]
+        for mount in (options.model, options.tokenizer):
+            if mount and Path(mount).exists():
+                cmd += ["-v", f"{mount}:{mount}:ro"]
         if options.dataset_path and Path(options.dataset_path).exists():
             cmd += ["-v", f"{options.dataset_path}:{options.dataset_path}:ro"]
         if options.output_file:
diff --git a/src/sskj/deploy/profile.py b/src/sskj/deploy/profile.py
index 077a08a..24d160a 100644
--- a/src/sskj/deploy/profile.py
+++ b/src/sskj/deploy/profile.py
@@ -57,13 +57,23 @@ def render_profile(profile: dict[str, str], overrides: dict[str, str] | None = N
         "PORT",
     ]
     rendered = dict(values)
-    for key in template_keys:
-        if key in rendered:
-            rendered[key] = expand_template(str(rendered[key]), values)
 
+    # Append DP_FLAG to LAUNCH_ARGS and BOOTSTRAP BEFORE template expansion.
+    # BOOTSTRAP's ${LAUNCH_ARGS} reference is already inlined during env-file
+    # parsing (without DP_FLAG), so appending to BOOTSTRAP directly is what
+    # guarantees the data-parallel flag reaches the actual launch command.
     dp = int(values["DP"])
     if dp > 1 and rendered.get("DP_FLAG"):
         rendered["LAUNCH_ARGS"] = f"{rendered.get('LAUNCH_ARGS', '')} {rendered['DP_FLAG']}".strip()
+        if "BOOTSTRAP" in rendered:
+            rendered["BOOTSTRAP"] = f"{rendered['BOOTSTRAP']} {rendered['DP_FLAG']}".strip()
+
+    # Expand against `rendered` (not `values`) so later references such as
+    # BOOTSTRAP's ${LAUNCH_ARGS} pick up earlier mutations (e.g. DP_FLAG).
+    for key in template_keys:
+        if key in rendered:
+            rendered[key] = expand_template(str(rendered[key]), rendered)
+
     if not rendered.get("RUNTIME"):
         rendered["RUNTIME"] = "docker" if rendered.get("DOCKER_IMAGE") else "native"
     return rendered