@@ -847,7 +847,7 @@ def _get_runner_type(self) -> str:
 
     def _prepare_benchmark_scripts(self) -> None:
         """
-        Copy Magpie generic benchmark scripts to InferenceX/benchmarks/.
+        Copy Magpie benchmark support files to InferenceX/benchmarks/.
         
         This allows using Magpie's generic scripts while still leveraging
         InferenceX's benchmark_lib.sh and other utilities.
@@ -865,11 +865,15 @@ def _prepare_benchmark_scripts(self) -> None:
         # Ensure target directory exists
         target_dir.mkdir(parents=True, exist_ok=True)
 
-        # Keep scripts in sync without replacing files used by concurrent runs.
-        for script in magpie_scripts.glob("*.sh"):
-            target_file = target_dir / script.name
-            self._copy_benchmark_script_atomic(script, target_file)
-            logger.info(f"Copied Magpie script {script.name} to {target_dir}")
+        # Keep support files in sync without replacing files used by concurrent runs.
+        support_files = [
+            *magpie_scripts.glob("*.sh"),
+            magpie_scripts / "lm_eval_humaneval_compat.py",
+        ]
+        for support_file in sorted(support_files):
+            target_file = target_dir / support_file.name
+            self._copy_benchmark_script_atomic(support_file, target_file)
+            logger.info(f"Copied Magpie support file {support_file.name} to {target_dir}")
 
     @staticmethod
     def _copy_benchmark_script_atomic(script: Path, target_file: Path) -> None:
 
@@ -0,0 +1,100 @@
+"""Run lm-eval with reliable assistant-prefix continuation for HumanEval."""
+
+from __future__ import annotations
+
+import runpy
+from collections.abc import Mapping, Sequence
+from typing import Any, Callable
+
+
+def open_final_assistant_continuation(
+    rendered: Any,
+    chat_history: Any,
+    *,
+    add_generation_prompt: bool,
+) -> Any:
+    """Remove template suffixes that incorrectly close a HumanEval prefix."""
+    if add_generation_prompt or not isinstance(rendered, str):
+        return rendered
+    if not isinstance(chat_history, Sequence) or not chat_history:
+        return rendered
+
+    final_message = chat_history[-1]
+    if not isinstance(final_message, Mapping):
+        return rendered
+    if final_message.get("role") != "assistant":
+        return rendered
+
+    prefix = final_message.get("content")
+    if not isinstance(prefix, str) or not prefix:
+        return rendered
+
+    prefix_start = rendered.rfind(prefix)
+    if prefix_start < 0:
+        return rendered
+
+    # Some tokenizer templates ignore continue_final_message and append their
+    # assistant/end-of-message markers. Preserve all model-native opening
+    # tokens while leaving the provided function prefix open for generation.
+    return rendered[: prefix_start + len(prefix)]
+
+
+def render_humaneval_chat_template(
+    model: Any,
+    chat_history: Any,
+    *,
+    add_generation_prompt: bool,
+    fallback: Callable[..., Any],
+) -> Any:
+    """Render a local HF template without changing the API prompt wire format."""
+    tokenizer = getattr(model, "tokenizer", None)
+    if getattr(model, "tokenizer_backend", None) == "huggingface" and tokenizer is not None:
+        rendered = tokenizer.apply_chat_template(
+            chat_history,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            continue_final_message=not add_generation_prompt,
+        )
+    else:
+        rendered = fallback(
+            model,
+            chat_history,
+            add_generation_prompt=add_generation_prompt,
+        )
+    return open_final_assistant_continuation(
+        rendered,
+        chat_history,
+        add_generation_prompt=add_generation_prompt,
+    )
+
+
+def install_humaneval_continuation_patch() -> None:
+    from lm_eval.models.api_models import TemplateAPI
+
+    original = TemplateAPI.apply_chat_template
+    if getattr(original, "_magpie_humaneval_continuation", False):
+        return
+
+    def apply_chat_template(
+        self,
+        chat_history: list[dict[str, str]],
+        add_generation_prompt: bool = True,
+    ) -> Any:
+        return render_humaneval_chat_template(
+            self,
+            chat_history,
+            add_generation_prompt=add_generation_prompt,
+            fallback=original,
+        )
+
+    apply_chat_template._magpie_humaneval_continuation = True  # type: ignore[attr-defined]
+    TemplateAPI.apply_chat_template = apply_chat_template
+
+
+def main() -> None:
+    install_humaneval_continuation_patch()
+    runpy.run_module("lm_eval.__main__", run_name="__main__", alter_sys=True)
+
+
+if __name__ == "__main__":
+    main()
@@ -413,24 +413,17 @@ magpie_eval_model_args() {
   local base_url="$2"
   local max_length="$3"
   local max_gen_toks="$4"
-  local args="model=${MODEL},base_url=${base_url},num_concurrent=${conc},tokenizer_backend=huggingface,trust_remote_code=true,max_length=${max_length},max_gen_toks=${max_gen_toks}"
+  local tokenizer="${MAGPIE_EVAL_TOKENIZER:-${MODEL}}"
+  local args="model=${MODEL},tokenizer=${tokenizer},base_url=${base_url},num_concurrent=${conc},tokenizer_backend=huggingface,trust_remote_code=true,max_length=${max_length},max_gen_toks=${max_gen_toks}"
+  if [[ -n "${MAGPIE_EVAL_TOKENIZER_REVISION:-}" ]]; then
+    args+=",revision=${MAGPIE_EVAL_TOKENIZER_REVISION}"
+  fi
   if [[ -n "${MAGPIE_EVAL_TOKENIZED_REQUESTS:-}" ]]; then
     args+=",tokenized_requests=${MAGPIE_EVAL_TOKENIZED_REQUESTS}"
   fi
   printf '%s\n' "$args"
 }
 
-magpie_eval_chat_model_args() {
-  local conc="$1"
-  local base_url="$2"
-  local max_length="$3"
-  local max_gen_toks="$4"
-  # Let the chat endpoint apply the model template. Otherwise lm-eval can send
-  # message objects as completion prompts, which compatible servers reject.
-  printf '%s\n' \
-    "model=${MODEL},base_url=${base_url},num_concurrent=${conc},tokenizer_backend=none,tokenized_requests=false,trust_remote_code=true,max_length=${max_length},max_gen_toks=${max_gen_toks}"
-}
-
 ###############################################################################
 # magpie_prepare_eval_include_and_limit
 #
@@ -608,8 +601,9 @@ PY
 #
 # Drive lm-eval with the local-completions backend so multiple-choice tasks
 # that need loglikelihood (mmlu, hellaswag) work against an OpenAI-compatible
-# /v1/completions server. InferenceX's local run_eval uses
-# local-chat-completions, which raises NotImplementedError for those requests.
+# /v1/completions server. HumanEval also uses that backend after lm-eval renders
+# its chat template locally, which preserves the task's assistant code prefix
+# as a continuation instead of sending it as an already completed chat message.
 #
 # Magpie builds the argv here so --confirm_run_unsafe_code is on the actual
 # lm-eval process even when a later InferenceX server-watch wrapper would
@@ -633,6 +627,18 @@ magpie_run_lm_eval_invocation() {
     echo "[magpie_bench_remote_compat] ERROR HumanEval tasks must run in a dedicated invocation" >&2
     return 1
   fi
+  local -a python_entrypoint=("$py" -m lm_eval)
+  if [[ "$humaneval_only" -eq 1 ]] && \
+     magpie_eval_truthy "${MAGPIE_EVAL_APPLY_CHAT_TEMPLATE:-}"; then
+    local compat_dir humaneval_runner
+    compat_dir=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P) || return 1
+    humaneval_runner="${MAGPIE_EVAL_HUMANEVAL_RUNNER:-${compat_dir}/lm_eval_humaneval_compat.py}"
+    if [[ ! -f "$humaneval_runner" ]]; then
+      echo "[magpie_bench_remote_compat] ERROR missing HumanEval runner: $humaneval_runner" >&2
+      return 1
+    fi
+    python_entrypoint=("$py" "$humaneval_runner")
+  fi
   local -a task_flag
   if [[ "$cli_mode" == "words" ]]; then
     task_flag=(--tasks "$@")
@@ -649,7 +655,7 @@ magpie_run_lm_eval_invocation() {
     task_flag=(--tasks "$joined")
   fi
   local -a cmd=(
-    "$py" -m lm_eval
+    "${python_entrypoint[@]}"
     --model "$model_backend"
     "${task_flag[@]}"
     --model_args "$model_args"
@@ -717,16 +723,6 @@ magpie_run_lm_eval() {
   local humaneval_model_backend="local-completions"
   local humaneval_model_args="$model_args"
   local humaneval_batch_size="$batch_size"
-  if magpie_eval_truthy "${MAGPIE_EVAL_APPLY_CHAT_TEMPLATE:-}"; then
-    local chat_base_url="${base_url%/completions}/chat/completions"
-    humaneval_model_backend="local-chat-completions"
-    humaneval_model_args="$(
-      magpie_eval_chat_model_args "$conc" "$chat_base_url" "$max_length" "$max_gen_toks"
-    )"
-    # Chat completions does not support request batching. num_concurrent still
-    # controls parallel requests through lm-eval's async client.
-    humaneval_batch_size="1"
-  fi
 
   local isolate_humaneval=0
   if magpie_eval_humaneval_options_enabled; then
@@ -836,6 +832,12 @@ magpie_run_lm_eval() {
 #   MAGPIE_EVAL_WRITE_OUT  true/1 to add --write_out to HumanEval
 #   MAGPIE_EVAL_APPLY_CHAT_TEMPLATE true/1 to add --apply_chat_template to
 #                         HumanEval. Mixed suites isolate HumanEval first.
+#   MAGPIE_EVAL_HUMANEVAL_RUNNER optional path to the HumanEval-only lm-eval
+#                         compatibility launcher.
+#   MAGPIE_EVAL_TOKENIZER Hugging Face tokenizer id or local path. Defaults to
+#                         MODEL; set it when MODEL is a serving alias or a
+#                         checkpoint whose tokenizer lives elsewhere.
+#   MAGPIE_EVAL_TOKENIZER_REVISION optional tokenizer repository revision.
 #
 # Returns lm-eval's exit code; prints diagnostics on stderr; never overrides
 # upstream lm-eval flags so future task adds are pure env tweaks.
 
@@ -4,6 +4,7 @@
 import subprocess
 import sys
 from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 
@@ -144,6 +145,10 @@ def _compat_script() -> Path:
     return ROOT / "Magpie" / "scripts" / "benchmark" / "magpie_bench_remote_compat.sh"
 
 
+def _humaneval_runner() -> Path:
+    return ROOT / "Magpie" / "scripts" / "benchmark" / "lm_eval_humaneval_compat.py"
+
+
 def _lm_eval_python_stub(tmp_path: Path, args_file: Path) -> Path:
     stub = tmp_path / "lm_eval_python"
     stub.write_text(
@@ -213,7 +218,7 @@ def test_remote_eval_uses_task_compatible_openai_backends():
         '${EVAL_CONCURRENT_REQUESTS:-${CONC:-8}}}"'
     ) in contents
     assert '"local-completions"' in contents
-    assert '"local-chat-completions"' in contents
+    assert "MAGPIE_EVAL_TOKENIZED_REQUESTS=true" not in contents
     assert "unset -f python3" not in contents
     assert "Crystal" not in contents
 
@@ -472,7 +477,9 @@ def test_remote_eval_adds_unsafe_flag_and_code_eval_env_for_humaneval(tmp_path:
     assert "HF_ALLOW_CODE_EVAL=1" in args
 
 
-def test_remote_eval_adds_humaneval_sample_logging_and_chat_template(tmp_path: Path):
+def test_remote_eval_adds_humaneval_sample_logging_and_local_chat_template(
+    tmp_path: Path,
+):
     args_file = tmp_path / "lm_eval.args"
     python_stub = _lm_eval_python_stub(tmp_path, args_file)
     shell = r'''
@@ -490,8 +497,10 @@ def test_remote_eval_adds_humaneval_sample_logging_and_chat_template(tmp_path: P
         "MAGPIE_EVAL_LOG_SAMPLES": "true",
         "MAGPIE_EVAL_WRITE_OUT": "true",
         "MAGPIE_EVAL_APPLY_CHAT_TEMPLATE": "true",
+        "MAGPIE_EVAL_TOKENIZER": "org/prompt-tokenizer",
+        "MAGPIE_EVAL_TOKENIZER_REVISION": "tokenizer-revision",
         "LM_EVAL_ARGS_FILE": str(args_file),
-        "MODEL": "test-model",
+        "MODEL": "served-model-alias",
     }
     env.pop("HF_ALLOW_CODE_EVAL", None)
     subprocess.run(["bash", "-c", shell], check=True, env=env)
@@ -501,11 +510,15 @@ def test_remote_eval_adds_humaneval_sample_logging_and_chat_template(tmp_path: P
     assert "--log_samples" in args
     assert "--write_out" in args
     assert "--apply_chat_template" in args
-    assert "--model local-chat-completions" in args
-    assert "base_url=http://127.0.0.1:8888/v1/chat/completions" in args
-    assert "tokenized_requests=false" in args
-    assert "tokenizer_backend=none" in args
-    assert "--batch_size 1" in args
+    assert "--model local-completions" in args
+    assert "base_url=http://127.0.0.1:8888/v1/completions" in args
+    assert "tokenized_requests=" not in args
+    assert "tokenizer_backend=huggingface" in args
+    assert "model=served-model-alias" in args
+    assert "tokenizer=org/prompt-tokenizer" in args
+    assert "revision=tokenizer-revision" in args
+    assert "--batch_size auto" in args
+    assert str(_humaneval_runner()) in args
 
 
 def test_remote_eval_splits_humaneval_flags_from_mixed_tasks(tmp_path: Path):
@@ -549,12 +562,20 @@ def test_remote_eval_splits_humaneval_flags_from_mixed_tasks(tmp_path: Path):
         assert flag in humaneval
         assert flag not in plain
         assert flag not in stock
+    assert str(_humaneval_runner()) in humaneval
+    assert str(_humaneval_runner()) not in plain
+    assert str(_humaneval_runner()) not in stock
+    assert "-m lm_eval" in plain
+    assert "-m lm_eval" in stock
     assert "--model local-completions" in plain
     assert "--model local-completions" in stock
-    assert "--model local-chat-completions" in humaneval
-    assert "base_url=http://127.0.0.1:8888/v1/chat/completions" in humaneval
+    assert "--model local-completions" in humaneval
+    assert "base_url=http://127.0.0.1:8888/v1/completions" in humaneval
+    assert "tokenized_requests=false" in humaneval
     assert "base_url=http://127.0.0.1:8888/v1/completions" in plain
+    assert "tokenized_requests=false" in plain
     assert "base_url=http://127.0.0.1:8888/v1/completions" in stock
+    assert "tokenized_requests=false" in stock
     assert "--include_path" not in plain
     assert "--include_path" not in humaneval
     assert f"--include_path {include_dir}" in stock
@@ -599,8 +620,52 @@ def test_remote_eval_preserves_stock_include_for_isolated_humaneval(tmp_path: Pa
         line for line in args_file.read_text().splitlines() if "--tasks" in line
     )
     assert "--tasks humaneval_instruct" in invocation
-    assert "--model local-chat-completions" in invocation
+    assert "--model local-completions" in invocation
+    assert "base_url=http://127.0.0.1:8888/v1/completions" in invocation
+    assert "tokenized_requests=false" in invocation
     assert f"--include_path {include_dir}" in invocation
+    assert str(_humaneval_runner()) in invocation
+
+
+def test_staged_humaneval_runner_is_invoked(tmp_path: Path):
+    from Magpie.modes.benchmark.benchmarker import BenchmarkMode
+
+    inferencex = tmp_path / "InferenceX"
+    mode = object.__new__(BenchmarkMode)
+    mode.config = SimpleNamespace(inferencex_path=str(inferencex))
+    mode._prepare_benchmark_scripts()
+
+    staged_dir = inferencex / "benchmarks"
+    staged_compat = staged_dir / "magpie_bench_remote_compat.sh"
+    staged_runner = staged_dir / "lm_eval_humaneval_compat.py"
+    assert staged_runner.read_bytes() == _humaneval_runner().read_bytes()
+
+    args_file = tmp_path / "lm_eval.args"
+    python_stub = _lm_eval_python_stub(tmp_path, args_file)
+    shell = r'''
+source "$MAGPIE_COMPAT"
+magpie_write_accuracy_result() { return 0; }
+magpie_run_eval_remote_direct
+'''
+    env = {
+        **os.environ,
+        "MAGPIE_COMPAT": str(staged_compat),
+        "RESULT_DIR": str(tmp_path / "result"),
+        "BENCHMARK_BASE_URL": "http://127.0.0.1:8888",
+        "MAGPIE_EVAL_PYTHON": str(python_stub),
+        "MAGPIE_EVAL_TASKS": "humaneval_instruct",
+        "MAGPIE_EVAL_APPLY_CHAT_TEMPLATE": "true",
+        "MAGPIE_EVAL_TOKENIZED_REQUESTS": "false",
+        "LM_EVAL_ARGS_FILE": str(args_file),
+        "MODEL": "test-model",
+    }
+    subprocess.run(["bash", "-c", shell], check=True, env=env)
+
+    invocation = next(
+        line for line in args_file.read_text().splitlines() if "--tasks" in line
+    )
+    assert str(staged_runner) in invocation
+    assert "tokenized_requests=false" in invocation
 
 
 def test_remote_eval_uses_comma_tasks_when_help_shows_single_string_cli(tmp_path: Path):