minor refactor

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
do not take &&, better error categorization
2026-07-24 09:43:05 +00:00 · 2026-07-24 09:26:14 +00:00 · 2026-07-24 09:22:01 +00:00 · 2026-07-24 09:10:46 +00:00 · 2026-07-23 10:16:32 +00:00
191 changed files with 2386 additions and 2799 deletions
@@ -1,11 +1,10 @@
 #!/bin/bash
 set -euox pipefail

-export VLLM_CPU_KVCACHE_SPACE=1
+export VLLM_CPU_KVCACHE_SPACE=1 
 export VLLM_CPU_CI_ENV=1
-# Skip torch.compile via vLLM's --enforce-eager flag (passed below) instead of
-# TORCH_COMPILE_DISABLE=1, which torch 2.12 no longer treats as a silent no-op
-# when callers specify fullgraph=True.
+# Reduce sub-processes for acceleration
+export TORCH_COMPILE_DISABLE=1 
 export VLLM_ENABLE_V1_MULTIPROCESSING=0

 SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
@@ -50,15 +49,15 @@ wait_for_pid_and_check_log() {
 }

 # Test Sky Lake (AVX512F)
-./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_0.log 2>&1 &
+./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
 PID_TEST_0=$!

 # Test Cascade Lake (AVX512F + VNNI)
-./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_1.log 2>&1 &
+./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
 PID_TEST_1=$!

 # Test Cooper Lake (AVX512F + VNNI + BF16)
-./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_2.log 2>&1 &
+./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
 PID_TEST_2=$!

 wait_for_pid_and_check_log $PID_TEST_0 test_0.log
@@ -5,7 +5,7 @@ steps:
 - label: PyTorch Compilation Unit Tests
  device: h200_35gb
  key: pytorch-compilation-unit-tests
-  timeout_in_minutes: 150
+  timeout_in_minutes: 110
  source_file_dependencies:
    - vllm/__init__.py
    - vllm/_aiter_ops.py
@@ -68,8 +68,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.13.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.13.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.11.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.11.0")
 # TORCH_NIGHTLY=1 builds run against unpinned nightly wheels, so the supported-
 # version check would always warn. Only treat it as a nightly build when the
 # value is exactly "1" (the bootstrap exports TORCH_NIGHTLY=0 by default, which
@@ -22,7 +22,7 @@
 #   docker buildx bake -f docker/docker-bake.hcl -f docker/versions.json
 # =============================================================================

-ARG CUDA_VERSION=13.0.3
+ARG CUDA_VERSION=13.0.2
 ARG PYTHON_VERSION=3.12
 ARG UBUNTU_VERSION=22.04

@@ -2,7 +2,7 @@
  "_comment": "Auto-generated from Dockerfile ARGs. Do not edit manually. Run: python tools/generate_versions_json.py",
  "variable": {
    "CUDA_VERSION": {
-      "default": "13.0.3"
+      "default": "13.0.2"
    },
    "PYTHON_VERSION": {
      "default": "3.12"
@@ -11,10 +11,10 @@
      "default": "22.04"
    },
    "BUILD_BASE_IMAGE": {
-      "default": "nvidia/cuda:13.0.3-devel-ubuntu22.04"
+      "default": "nvidia/cuda:13.0.2-devel-ubuntu22.04"
    },
    "FINAL_BASE_IMAGE": {
-      "default": "nvidia/cuda:13.0.3-base-ubuntu22.04"
+      "default": "nvidia/cuda:13.0.2-base-ubuntu22.04"
    },
    "BUILD_OS": {
      "default": "ubuntu"
@@ -7,7 +7,7 @@ requires = [
    "setuptools>=77.0.3,<81.0.0",
    "setuptools-scm>=8.0",
    "setuptools-rust>=1.9.0",
-    "torch == 2.13.0",
+    "torch == 2.11.0",
    "wheel",
    "jinja2",
 ]
@@ -4,8 +4,8 @@ packaging>=24.2
 setuptools==77.0.3 # this version can reuse CMake build dir
 setuptools-scm>=8
 setuptools-rust>=1.9.0
-torch==2.13.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"
-torch==2.13.0; platform_system == "Darwin" or platform_machine == "ppc64le"  or platform_machine == "riscv64"
+torch==2.11.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"
+torch==2.11.0; platform_system == "Darwin" or platform_machine == "ppc64le"  or platform_machine == "riscv64"
 wheel
 jinja2>=3.1.6
 regex
@@ -5,7 +5,7 @@ packaging>=24.2
 setuptools>=77.0.3,<81.0.0
 setuptools-scm>=8
 setuptools-rust>=1.9.0
-torch==2.13.0
+torch==2.11.0
 wheel
 jinja2>=3.1.6
 regex
@@ -6,8 +6,8 @@ setuptools==77.0.3 # this version can reuse CMake build dir
 numba == 0.65.0; platform_machine != "s390x" # Required for N-gram speculative decoding

 # Dependencies for CPUs
-torch==2.13.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"
-torch==2.13.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64"
+torch==2.11.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"
+torch==2.11.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64"

 # required for the image processor of minicpm-o-2_6, this must be updated alongside torch
 torchaudio; platform_machine != "s390x" and platform_machine != "riscv64"
@@ -4,10 +4,10 @@
 numba == 0.65.0 # Required for N-gram speculative decoding

 # Dependencies for NVIDIA GPUs
-torch==2.13.0
+torch==2.11.0
 torchaudio==2.11.0
 # These must be updated alongside torch
-torchvision==0.28.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
+torchvision==0.26.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
 torchcodec >= 0.14
 PyNvVideoCodec==2.0.4
 # FlashInfer should be updated together with the Dockerfile
@@ -1107,7 +1107,7 @@ tokenizers==0.22.2
    #   -r requirements/test/../common.txt
    #   -r requirements/test/cuda.in
    #   transformers
-torch==2.13.0+cpu
+torch==2.11.0+cpu
    # via
    #   -r requirements/test/cuda.in
    #   accelerate
@@ -1134,7 +1134,7 @@ torchaudio==2.11.0+cpu
    #   vocos
 torchcodec==0.14.0+cpu
    # via -r requirements/test/cuda.in
-torchvision==0.28.0+cpu
+torchvision==0.26.0+cpu
    # via
    #   -r requirements/test/cuda.in
    #   open-clip-torch
@@ -28,9 +28,9 @@ soundfile # required for audio tests
 jiwer # required for audio tests
 tblib # for pickling test exceptions
 timm >=1.0.17 # required for internvl and gemma3n-mm test
-torch==2.13.0
+torch==2.11.0
 torchaudio==2.11.0
-torchvision==0.28.0
+torchvision==0.26.0
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
 mistral_common[image,audio] >= 1.11.5 # required for voxtral test
@@ -159,7 +159,7 @@ cuda-bindings==13.0.3
    # via torch
 cuda-pathfinder==1.3.3
    # via cuda-bindings
-cuda-toolkit==13.0.3.0
+cuda-toolkit==13.0.2
    # via torch
 cupy-cuda12x==13.6.0
    # via ray
@@ -599,7 +599,7 @@ numpy==2.2.6
    #   tritonclient
    #   vocos
    #   xgrammar
-nvidia-cublas==13.1.1.3
+nvidia-cublas==13.1.0.3
    # via
    #   cuda-toolkit
    #   nvidia-cudnn-cu13
@@ -607,12 +607,10 @@ nvidia-cublas==13.1.1.3
 nvidia-cuda-cupti==13.0.85
    # via cuda-toolkit
 nvidia-cuda-nvrtc==13.0.88
-    # via
-    #   cuda-toolkit
-    #   nvidia-cublas
+    # via cuda-toolkit
 nvidia-cuda-runtime==13.0.96
    # via cuda-toolkit
-nvidia-cudnn-cu13==9.20.0.48
+nvidia-cudnn-cu13==9.19.0.56
    # via torch
 nvidia-cufft==12.0.0.61
    # via cuda-toolkit
@@ -626,9 +624,9 @@ nvidia-cusparse==12.6.3.3
    # via
    #   cuda-toolkit
    #   nvidia-cusolver
-nvidia-cusparselt-cu13==0.8.1
+nvidia-cusparselt-cu13==0.8.0
    # via torch
-nvidia-nccl-cu13==2.29.7
+nvidia-nccl-cu13==2.28.9
    # via torch
 nvidia-nvjitlink==13.0.88
    # via
@@ -1204,7 +1202,7 @@ tokenizers==0.22.2
    #   -r requirements/test/../common.txt
    #   -r requirements/test/cuda.in
    #   transformers
-torch==2.13.0+cu130
+torch==2.11.0+cu130
    # via
    #   -c requirements/cuda.txt
    #   -r requirements/test/cuda.in
@@ -1235,7 +1233,7 @@ torchcodec==0.14.0+cu130
    # via
    #   -c requirements/cuda.txt
    #   -r requirements/test/cuda.in
-torchvision==0.28.0+cu130
+torchvision==0.26.0+cu130
    # via
    #   -c requirements/cuda.txt
    #   -r requirements/test/cuda.in
@@ -1272,7 +1270,7 @@ transformers==5.13.1
    #   xgrammar
 transformers-stream-generator==0.0.5
    # via -r requirements/test/cuda.in
-triton==3.7.1
+triton==3.6.0
    # via
    #   torch
    #   xgrammar
@@ -4252,18 +4252,6 @@ dependencies = [
 "syn 2.0.117",
 ]

-[[package]]
-name = "subenum"
-version = "1.1.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ec3d08fe7078c57309d5c3d938e50eba95ba1d33b9c3a101a8465fc6861a5416"
-dependencies = [
- "heck",
- "proc-macro2",
- "quote",
- "syn 2.0.117",
-]
-
 [[package]]
 name = "subtle"
 version = "2.6.1"
@@ -5516,15 +5504,11 @@ dependencies = [
 "asynk-strim-attr",
 "bytes",
 "clap",
- "easy-ext",
 "expect-test",
 "futures",
 "half",
- "indexmap 2.13.0",
 "itertools 0.14.0",
 "llm-multimodal",
- "minijinja",
- "minijinja-contrib",
 "ndarray 0.17.2",
 "oss-harmony",
 "paste",
@@ -5535,17 +5519,16 @@ dependencies = [
 "serde_json",
 "serde_with",
 "serial_test",
- "strum",
- "subenum",
 "tempfile",
 "thiserror 2.0.18",
 "thiserror-ext",
- "time",
 "tokio",
 "tracing",
 "tracing-subscriber",
 "trait-set",
 "uuid",
+ "vllm-chat-renderer",
+ "vllm-chat-types",
 "vllm-engine-core-client",
 "vllm-llm",
 "vllm-parser",
@@ -5555,6 +5538,42 @@ dependencies = [
 "zeromq",
 ]

+[[package]]
+name = "vllm-chat-renderer"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "enum-as-inner",
+ "expect-test",
+ "indexmap 2.13.0",
+ "itertools 0.14.0",
+ "minijinja",
+ "minijinja-contrib",
+ "oss-harmony",
+ "serde",
+ "serde-json-fmt",
+ "serde_json",
+ "serde_with",
+ "strum",
+ "tempfile",
+ "thiserror 2.0.18",
+ "thiserror-ext",
+ "time",
+ "tracing",
+ "vllm-chat-types",
+ "vllm-tokenizer",
+]
+
+[[package]]
+name = "vllm-chat-types"
+version = "0.1.0"
+dependencies = [
+ "easy-ext",
+ "serde",
+ "serde_json",
+ "serde_with",
+]
+
 [[package]]
 name = "vllm-cmd"
 version = "0.1.0"
@@ -5695,11 +5714,11 @@ dependencies = [
 "expect-test",
 "futures",
 "openai-protocol",
- "serde",
 "serde_json",
 "thiserror 2.0.18",
 "thiserror-ext",
 "tool-parser",
+ "vllm-chat-types",
 "vllm-tokenizer",
 "winnow",
 "xgrammar-structural-tag",
@@ -5806,6 +5825,7 @@ dependencies = [
 "rustc-hash 1.1.0",
 "serde",
 "serde_json",
+ "serde_with",
 "tekken-rs",
 "tempfile",
 "thiserror 2.0.18",
@@ -2,6 +2,8 @@
 members = [
    "src/bench",
    "src/chat",
+    "src/chat-renderer",
+    "src/chat-types",
    "src/cmd",
    "src/engine-core-client",
    "src/llm",
@@ -135,6 +137,8 @@ uuid = { version = "1.22.0", features = ["v4"] }
 validator = { version = "0.20.0", features = ["derive"] }
 vllm-bench = { path = "src/bench" }
 vllm-chat = { path = "src/chat" }
+vllm-chat-renderer = { path = "src/chat-renderer" }
+vllm-chat-types = { path = "src/chat-types" }
 vllm-engine-core-client = { path = "src/engine-core-client" }
 vllm-llm = { path = "src/llm" }
 vllm-managed-engine = { path = "src/managed-engine" }
@@ -0,0 +1,32 @@
+[package]
+name = "vllm-chat-renderer"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+enum-as-inner.workspace = true
+indexmap.workspace = true
+itertools.workspace = true
+minijinja.workspace = true
+minijinja-contrib.workspace = true
+openai-harmony.workspace = true
+serde.workspace = true
+serde-json-fmt.workspace = true
+serde_json.workspace = true
+serde_with.workspace = true
+strum.workspace = true
+thiserror.workspace = true
+thiserror-ext.workspace = true
+time.workspace = true
+tracing.workspace = true
+vllm-chat-types.workspace = true
+vllm-tokenizer.workspace = true
+
+[dev-dependencies]
+expect-test.workspace = true
+tempfile.workspace = true
+
+[lints]
+workspace = true
@@ -11,8 +11,10 @@ use serde_json::Value;
 use serde_json_fmt::JsonFormat;

 use crate::error::{Error, Result};
-use crate::request::{ChatContent, ChatMessage, ChatRequest, ChatRole, ChatTool};
-use crate::{AssistantContentBlock, AssistantMessageExt, AssistantToolCall};
+use crate::{
+    AssistantContentBlock, AssistantMessageExt, AssistantToolCall, ChatContent, ChatMessage,
+    ChatRole, RenderRequest, Tool,
+};

 const BOS_TOKEN: &str = "<｜begin▁of▁sentence｜>";
 const EOS_TOKEN: &str = "<｜end▁of▁sentence｜>";
@@ -39,7 +41,7 @@ struct RenderedToolSchema<'a> {
 }

 /// Render one chat request into the final prompt string.
-pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
+pub(super) fn render_request(request: &RenderRequest<'_>) -> Result<String> {
    let thinking_mode = match request.enable_thinking()?.unwrap_or(false) {
        true => ThinkingMode::Thinking,
        false => ThinkingMode::Chat,
@@ -49,20 +51,19 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
        Some(ChatRole::User | ChatRole::Developer)
    );
    let render_offset = isize::from(request.tool_parsing_enabled());
-    let last_user_render_index =
-        find_last_user_render_index(request.messages.as_slice(), render_offset);
-    let last_user_actual_index = find_last_user_actual_index(request.messages.as_slice());
+    let last_user_render_index = find_last_user_render_index(request.messages, render_offset);
+    let last_user_actual_index = find_last_user_actual_index(request.messages);
    let continue_final_message = request.chat_options.continue_final_message();
    let mut prompt = String::from(BOS_TOKEN);

    if request.tool_parsing_enabled() {
-        render_system_message(&mut prompt, None, &request.tools)?;
+        render_system_message(&mut prompt, None, request.tools)?;
    }

    for (message_index, message) in request.messages.iter().enumerate() {
        render_message(
            &mut prompt,
-            request.messages.as_slice(),
+            request.messages,
            message_index,
            message,
            render_offset,
@@ -168,7 +169,7 @@ fn find_last_user_actual_index(messages: &[ChatMessage]) -> usize {
 fn render_system_message(
    out: &mut String,
    content: Option<&ChatContent>,
-    tools: &[ChatTool],
+    tools: &[Tool],
 ) -> Result<()> {
    if let Some(content) = content {
        write_chat_content(out, content)?;
@@ -185,7 +186,7 @@ fn render_system_message(
 fn render_developer_message(
    out: &mut String,
    content: &ChatContent,
-    tools: &[ChatTool],
+    tools: &[Tool],
    opens_thinking: bool,
 ) -> Result<()> {
    if content.is_empty() {
@@ -452,7 +453,7 @@ fn encode_arguments_to_dsml(out: &mut String, tool_call: &AssistantToolCall) ->
 }

 /// Render the full tool preamble shown to the model.
-fn render_tools(out: &mut String, tools: &[ChatTool]) -> Result<()> {
+fn render_tools(out: &mut String, tools: &[Tool]) -> Result<()> {
    out.push_str(
        r#"## Tools

@@ -500,7 +501,7 @@ Here are the functions available in JSONSchema format:

 /// Serialize one typed tool schema into the JSON shape embedded inside
 /// `<functions>`.
-fn render_tool_schema(out: &mut String, tool: &ChatTool) -> Result<()> {
+fn render_tool_schema(out: &mut String, tool: &Tool) -> Result<()> {
    out.push_str(&json_dumps(&RenderedToolSchema {
        name: &tool.name,
        description: tool.description.as_deref(),
@@ -517,7 +518,7 @@ fn write_chat_content(out: &mut String, content: &ChatContent) -> Result<()> {
        ChatContent::Text(text) => out.push_str(text),
        ChatContent::Parts(parts) => {
            for part in parts {
-                out.push_str(part.as_text()?);
+                out.push_str(part.as_text().map_err(Error::UnsupportedMultimodalContent)?);
            }
        }
    }
@@ -3,11 +3,10 @@

 mod encoding;

-use vllm_text::Prompt;
-
-use super::{ChatRenderer, RenderedPrompt, request_template_kwargs};
+use super::{
+    ChatRenderer, RenderRequest, RenderedPrompt, RenderedPromptContent, request_template_kwargs,
+};
 use crate::Result;
-use crate::request::ChatRequest;

 /// Dedicated DeepSeek V3.2 renderer.
 #[derive(Debug, Clone, Copy, Default)]
@@ -21,12 +20,12 @@ impl DeepSeekV32ChatRenderer {
 }

 impl ChatRenderer for DeepSeekV32ChatRenderer {
-    fn render(&self, request: &ChatRequest) -> Result<RenderedPrompt> {
+    fn render(&self, request: RenderRequest<'_>) -> Result<RenderedPrompt> {
        request.validate()?;

        Ok(RenderedPrompt {
-            prompt: Prompt::Text(encoding::render_request(request)?),
-            effective_template_kwargs: request_template_kwargs(request),
+            content: RenderedPromptContent::Text(encoding::render_request(&request)?),
+            effective_template_kwargs: request_template_kwargs(&request),
        })
    }
 }
@@ -9,36 +9,35 @@ use thiserror_ext::AsReport;

 use super::DeepSeekV32ChatRenderer;
 use crate::error::Error;
-use crate::event::{AssistantContentBlock, AssistantToolCall};
-use crate::renderer::test_utils::{FixtureRequestOptions, fixture_chat_request};
-use crate::request::{
-    ChatContentPart, ChatMessage, ChatRequest, ChatTool, ChatToolChoice, GenerationPromptMode,
+use crate::test_utils::{FixtureRequestOptions, fixture_chat_request};
+use crate::{AssistantContentBlock, AssistantToolCall};
+use crate::{
+    ChatContentPart, ChatMessage, ChatToolChoice, GenerationPromptMode, TestRenderRequest, Tool,
 };
 use crate::{ChatRenderer, ChatRole};

-fn render_request(request: &ChatRequest) -> String {
+fn render_request(request: &TestRenderRequest) -> String {
    DeepSeekV32ChatRenderer::new()
-        .render(request)
+        .render(request.as_request())
        .unwrap()
-        .prompt
+        .content
        .into_text()
        .expect("deepseek renderer should return text prompt")
 }

-fn render_result(request: &ChatRequest) -> Result<String, Error> {
-    DeepSeekV32ChatRenderer::new().render(request).map(|rendered| {
+fn render_result(request: &TestRenderRequest) -> Result<String, Error> {
+    DeepSeekV32ChatRenderer::new().render(request.as_request()).map(|rendered| {
        rendered
-            .prompt
+            .content
            .into_text()
            .expect("deepseek renderer should return text prompt")
    })
 }

-fn thinking_request(messages: Vec<ChatMessage>) -> ChatRequest {
-    let mut request = ChatRequest {
-        request_id: "deepseek-v32-small-test".to_string(),
+fn thinking_request(messages: Vec<ChatMessage>) -> TestRenderRequest {
+    let mut request = TestRenderRequest {
        messages,
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    if matches!(
        request.messages.last().map(ChatMessage::role),
@@ -53,7 +52,7 @@ fn thinking_request(messages: Vec<ChatMessage>) -> ChatRequest {
    request
 }

-fn fixture_request(input_name: &str) -> ChatRequest {
+fn fixture_request(input_name: &str) -> TestRenderRequest {
    fixture_chat_request(&fixture_path(input_name), deepseek_fixture_options())
 }

@@ -66,7 +65,7 @@ fn deepseek_fixture_options() -> FixtureRequestOptions {

 fn fixture_path(name: &str) -> PathBuf {
    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
-        .join("src/renderer/deepseek_v32")
+        .join("src/deepseek_v32")
        .join("fixtures")
        .join(name)
 }
@@ -103,13 +102,12 @@ fn renders_official_search_fixture_with_date() {

 #[test]
 fn request_level_tools_are_lowered_as_synthetic_leading_system_message() {
-    let mut request = ChatRequest {
-        request_id: "deepseek-v32-tools".to_string(),
+    let mut request = TestRenderRequest {
        messages: vec![
            ChatMessage::system("System prompt."),
            ChatMessage::text(ChatRole::User, "Hello"),
        ],
-        tools: vec![ChatTool {
+        tools: vec![Tool {
            name: "lookup".to_string(),
            description: Some("Look things up".to_string()),
            parameters: json!({
@@ -124,7 +122,7 @@ fn request_level_tools_are_lowered_as_synthetic_leading_system_message() {
            strict: None,
        }],
        tool_choice: ChatToolChoice::Auto,
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request
        .chat_options
@@ -272,12 +270,12 @@ fn assistant_after_last_user_requires_reasoning_or_tool_calls() {

 #[test]
 fn continue_final_assistant_omits_final_eos() {
-    let mut request = ChatRequest {
+    let mut request = TestRenderRequest {
        messages: vec![
            ChatMessage::user("write"),
            ChatMessage::assistant_text("partial answer"),
        ],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request.chat_options.generation_prompt_mode = GenerationPromptMode::ContinueFinalAssistant;

@@ -289,14 +287,14 @@ fn continue_final_assistant_omits_final_eos() {

 #[test]
 fn render_rejects_multimodal_input() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![ChatMessage::user(vec![ChatContentPart::image_url(
            "data:image/png;base64,test",
        )])],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

-    let error = DeepSeekV32ChatRenderer::new().render(&request).unwrap_err();
+    let error = DeepSeekV32ChatRenderer::new().render(request.as_request()).unwrap_err();

    assert!(matches!(
        error,
@@ -14,8 +14,10 @@ use serde_json::Value;
 use serde_json_fmt::JsonFormat;

 use crate::error::{Error, Result};
-use crate::request::{ChatContent, ChatMessage, ChatRequest, ChatTool, ReasoningEffort};
-use crate::{AssistantContentBlock, AssistantMessageExt, AssistantToolCall};
+use crate::{
+    AssistantContentBlock, AssistantMessageExt, AssistantToolCall, ChatContent, ChatMessage,
+    ReasoningEffort, RenderRequest, Tool,
+};

 const BOS_TOKEN: &str = "<｜begin▁of▁sentence｜>";
 const EOS_TOKEN: &str = "<｜end▁of▁sentence｜>";
@@ -46,14 +48,14 @@ struct RenderedToolSchema<'a> {
 }

 /// Render one chat request into the final prompt string.
-pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
+pub(super) fn render_request(request: &RenderRequest<'_>) -> Result<String> {
    let (thinking_mode, max_reasoning_effort) = resolve_thinking_options(request)?;
    let request_tools = request_tools(request);
    let synthetic_tool_system = needs_synthetic_tool_system(request, request_tools);
    let drop_thinking = request.parse_template_bool("drop_thinking")?.unwrap_or(true)
        && !rendered_tools_present(request, request_tools);
    let last_user_render_index =
-        find_last_user_render_index(request.messages.as_slice(), synthetic_tool_system);
+        find_last_user_render_index(request.messages, synthetic_tool_system);
    let mut out = String::from(BOS_TOKEN);
    if thinking_mode == ThinkingMode::Thinking && max_reasoning_effort {
        out.push_str(REASONING_EFFORT_MAX);
@@ -68,7 +70,7 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
    }

    for (message_index, message) in request.messages.iter().enumerate() {
-        if is_following_tool_response(request.messages.as_slice(), message_index) {
+        if is_following_tool_response(request.messages, message_index) {
            continue;
        }

@@ -101,12 +103,12 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
                render_assistant_message(&mut out, emit_thinking_block, append_eos, content)?;
            }
            ChatMessage::ToolResponse { .. } => {
-                render_tool_response_block(&mut out, request.messages.as_slice(), message_index)?;
+                render_tool_response_block(&mut out, request.messages, message_index)?;
            }
        }

        if is_user_like_entry(message)
-            && next_rendered_entry_is_assistant_or_end(request.messages.as_slice(), message_index)
+            && next_rendered_entry_is_assistant_or_end(request.messages, message_index)
        {
            write_assistant_transition(
                &mut out,
@@ -124,7 +126,7 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
 /// wrapper, the Rust renderer only consumes the typed top-level
 /// `reasoning_effort`; the generic template-kwargs map is left for HF
 /// templates.
-fn resolve_thinking_options(request: &ChatRequest) -> Result<(ThinkingMode, bool)> {
+fn resolve_thinking_options(request: &RenderRequest<'_>) -> Result<(ThinkingMode, bool)> {
    let mut thinking_mode = match request.enable_thinking()?.unwrap_or(false) {
        true => ThinkingMode::Thinking,
        false => ThinkingMode::Chat,
@@ -141,16 +143,16 @@ fn resolve_thinking_options(request: &ChatRequest) -> Result<(ThinkingMode, bool
 }

 /// Return request-level tools only when native tool parsing is enabled.
-fn request_tools(request: &ChatRequest) -> &[ChatTool] {
+fn request_tools<'a>(request: &RenderRequest<'a>) -> &'a [Tool] {
    if request.tool_parsing_enabled() {
-        request.tools.as_slice()
+        request.tools
    } else {
        &[]
    }
 }

 /// Return whether request tools need a synthetic leading system entry.
-fn needs_synthetic_tool_system(request: &ChatRequest, request_tools: &[ChatTool]) -> bool {
+fn needs_synthetic_tool_system(request: &RenderRequest<'_>, request_tools: &[Tool]) -> bool {
    !request_tools.is_empty()
        && !request
            .messages
@@ -159,7 +161,7 @@ fn needs_synthetic_tool_system(request: &ChatRequest, request_tools: &[ChatTool]
 }

 /// Return whether any rendered message carries tool schemas.
-fn rendered_tools_present(request: &ChatRequest, request_tools: &[ChatTool]) -> bool {
+fn rendered_tools_present(request: &RenderRequest<'_>, request_tools: &[Tool]) -> bool {
    !request_tools.is_empty()
        || request.messages.iter().any(|message| {
            matches!(
@@ -228,7 +230,7 @@ fn next_rendered_entry_is_assistant_or_end(messages: &[ChatMessage], message_ind
 }

 /// Render the tool preamble shown to the model, V4 flavor.
-fn render_tools(out: &mut String, tools: &[ChatTool]) -> Result<()> {
+fn render_tools(out: &mut String, tools: &[Tool]) -> Result<()> {
    out.push_str(
        r#"## Tools

@@ -269,7 +271,7 @@ Otherwise, output directly after </think> with tool calls or final response.
 }

 /// Serialize one typed tool schema into the JSON shape embedded in the prompt.
-fn render_tool_schema(out: &mut String, tool: &ChatTool) -> Result<()> {
+fn render_tool_schema(out: &mut String, tool: &Tool) -> Result<()> {
    out.push_str(&json_dumps(&RenderedToolSchema {
        name: &tool.name,
        description: tool.description.as_deref(),
@@ -283,7 +285,7 @@ fn render_tool_schema(out: &mut String, tool: &ChatTool) -> Result<()> {
 fn render_system_message(
    out: &mut String,
    content: Option<&ChatContent>,
-    tools: &[ChatTool],
+    tools: &[Tool],
 ) -> Result<()> {
    if let Some(content) = content {
        write_chat_content(out, content)?;
@@ -296,11 +298,7 @@ fn render_system_message(
 }

 /// Developer messages are rendered as user-like turns with optional tools.
-fn render_developer_message(
-    out: &mut String,
-    content: &ChatContent,
-    tools: &[ChatTool],
-) -> Result<()> {
+fn render_developer_message(out: &mut String, content: &ChatContent, tools: &[Tool]) -> Result<()> {
    if content.is_empty() {
        return Err(Error::ChatTemplate(
            "invalid DeepSeek V4 developer message: empty content".to_string(),
@@ -519,7 +517,7 @@ fn write_chat_content(out: &mut String, content: &ChatContent) -> Result<()> {
        ChatContent::Text(text) => out.push_str(text),
        ChatContent::Parts(parts) => {
            for part in parts {
-                out.push_str(part.as_text()?);
+                out.push_str(part.as_text().map_err(Error::UnsupportedMultimodalContent)?);
            }
        }
    }
@@ -3,29 +3,29 @@

 mod encoding;

-use vllm_text::Prompt;
-
-use super::{ChatRenderer, RenderedPrompt, request_template_kwargs};
+use super::{
+    ChatRenderer, RenderRequest, RenderedPrompt, RenderedPromptContent, request_template_kwargs,
+};
 use crate::Result;
-use crate::request::ChatRequest;

 /// Dedicated DeepSeek V4 renderer.
 #[derive(Debug, Clone, Copy, Default)]
 pub struct DeepSeekV4ChatRenderer;

 impl DeepSeekV4ChatRenderer {
+    /// Create the dedicated DeepSeek V4 renderer.
    pub fn new() -> Self {
        Self
    }
 }

 impl ChatRenderer for DeepSeekV4ChatRenderer {
-    fn render(&self, request: &ChatRequest) -> Result<RenderedPrompt> {
+    fn render(&self, request: RenderRequest<'_>) -> Result<RenderedPrompt> {
        request.validate()?;

        Ok(RenderedPrompt {
-            prompt: Prompt::Text(encoding::render_request(request)?),
-            effective_template_kwargs: request_template_kwargs(request),
+            content: RenderedPromptContent::Text(encoding::render_request(&request)?),
+            effective_template_kwargs: request_template_kwargs(&request),
        })
    }
 }
@@ -8,20 +8,20 @@ use serde_json::Value;

 use super::DeepSeekV4ChatRenderer;
 use crate::ChatRenderer;
-use crate::event::{AssistantContentBlock, AssistantToolCall};
-use crate::renderer::test_utils::{FixtureRequestOptions, fixture_chat_request};
-use crate::request::{ChatMessage, ChatRequest, GenerationPromptMode, ReasoningEffort};
+use crate::test_utils::{FixtureRequestOptions, fixture_chat_request};
+use crate::{AssistantContentBlock, AssistantToolCall};
+use crate::{ChatMessage, GenerationPromptMode, ReasoningEffort, TestRenderRequest};

-fn render_request(request: &ChatRequest) -> String {
+fn render_request(request: &TestRenderRequest) -> String {
    DeepSeekV4ChatRenderer::new()
-        .render(request)
+        .render(request.as_request())
        .unwrap()
-        .prompt
+        .content
        .into_text()
        .expect("deepseek v4 renderer should return text prompt")
 }

-fn fixture_request(input_name: &str) -> ChatRequest {
+fn fixture_request(input_name: &str) -> TestRenderRequest {
    fixture_chat_request(&fixture_path(input_name), deepseek_fixture_options())
 }

@@ -34,7 +34,7 @@ fn deepseek_fixture_options() -> FixtureRequestOptions {

 fn fixture_path(name: &str) -> PathBuf {
    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
-        .join("src/renderer/deepseek_v4")
+        .join("src/deepseek_v4")
        .join("fixtures")
        .join(name)
 }
@@ -63,9 +63,9 @@ fn renders_v4_fixture_2_multi_turn_drop_thinking() {

 #[test]
 fn reasoning_effort_max_adds_prefix_when_thinking_is_enabled() {
-    let mut request = ChatRequest {
+    let mut request = TestRenderRequest {
        messages: vec![ChatMessage::user("solve it")],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request
        .chat_options
@@ -86,9 +86,9 @@ fn reasoning_effort_max_adds_prefix_when_thinking_is_enabled() {

 #[test]
 fn reasoning_effort_none_disables_thinking() {
-    let mut request = ChatRequest {
+    let mut request = TestRenderRequest {
        messages: vec![ChatMessage::user("answer directly")],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request
        .chat_options
@@ -104,9 +104,9 @@ fn reasoning_effort_none_disables_thinking() {

 #[test]
 fn reasoning_effort_template_kwarg_is_ignored() {
-    let mut request = ChatRequest {
+    let mut request = TestRenderRequest {
        messages: vec![ChatMessage::user("solve it")],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request
        .chat_options
@@ -124,7 +124,7 @@ fn reasoning_effort_template_kwarg_is_ignored() {

 #[test]
 fn tool_results_are_sorted_by_previous_assistant_tool_call_order() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![
            ChatMessage::assistant_blocks(vec![
                AssistantContentBlock::ToolCall(AssistantToolCall {
@@ -141,7 +141,7 @@ fn tool_results_are_sorted_by_previous_assistant_tool_call_order() {
            ChatMessage::tool_response("first result", "first"),
            ChatMessage::tool_response("second result", "second"),
        ],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

    let rendered = render_request(&request);
@@ -164,7 +164,7 @@ fn tool_results_are_sorted_by_previous_assistant_tool_call_order() {

 #[test]
 fn drop_thinking_false_keeps_prior_assistant_reasoning() {
-    let mut request = ChatRequest {
+    let mut request = TestRenderRequest {
        messages: vec![
            ChatMessage::assistant_blocks(vec![
                AssistantContentBlock::Reasoning {
@@ -176,7 +176,7 @@ fn drop_thinking_false_keeps_prior_assistant_reasoning() {
            ]),
            ChatMessage::user("next"),
        ],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request
        .chat_options
@@ -197,16 +197,16 @@ fn drop_thinking_false_keeps_prior_assistant_reasoning() {

 #[test]
 fn continue_final_assistant_omits_final_eos() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![
            ChatMessage::user("write"),
            ChatMessage::assistant_text("partial answer"),
        ],
-        chat_options: crate::request::ChatOptions {
+        chat_options: crate::ChatOptions {
            generation_prompt_mode: GenerationPromptMode::ContinueFinalAssistant,
            ..Default::default()
        },
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

    let rendered = render_request(&request);
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+use thiserror::Error;
+
+type BoxedError = Box<dyn std::error::Error + Send + Sync>;
+
+/// Error returned while constructing or applying a chat renderer.
+#[derive(Debug, Error)]
+pub enum Error {
+    /// Rendering requires at least one chat message.
+    #[error("chat request must contain at least one message")]
+    EmptyMessages,
+    /// Continuation mode requires the final message to be an assistant turn.
+    #[error("cannot continue the final message when the last message is not from the assistant")]
+    ContinueFinalAssistantWithoutFinalAssistant,
+    /// The selected renderer requires a chat template and none was configured.
+    #[error("chat template is required but none was configured")]
+    MissingChatTemplate,
+    /// A chat template could not be compiled or applied.
+    #[error("chat template error: {0}")]
+    ChatTemplate(String),
+    /// The selected renderer cannot represent the given multimodal part.
+    #[error("unsupported multimodal content: {0}")]
+    UnsupportedMultimodalContent(&'static str),
+    /// The process-wide GPT-OSS Harmony encoding could not be initialized.
+    #[error("failed to initialize the Harmony encoding")]
+    HarmonyEncoding {
+        /// Underlying Harmony initialization failure.
+        #[source]
+        error: BoxedError,
+    },
+    /// Tokenizer construction or encoding failed.
+    #[error(transparent)]
+    Tokenizer(#[from] vllm_tokenizer::TokenizerError),
+}
+
+impl Error {
+    /// Whether this error should be reported as invalid request input when
+    /// raised while rendering.
+    pub fn is_request_validation_error(&self) -> bool {
+        matches!(
+            self,
+            Self::EmptyMessages
+                | Self::ContinueFinalAssistantWithoutFinalAssistant
+                | Self::MissingChatTemplate
+                | Self::ChatTemplate(_)
+                | Self::UnsupportedMultimodalContent(_)
+        )
+    }
+}
+
+/// Result returned by chat renderer operations.
+pub type Result<T> = std::result::Result<T, Error>;
@@ -12,13 +12,13 @@ use thiserror_ext::AsReport as _;
 use crate::error::{Error, Result};

 /// Lazily load the shared GPT-OSS Harmony encoding once per process.
-pub(crate) fn harmony_encoding() -> Result<&'static HarmonyEncoding> {
+pub fn harmony_encoding() -> Result<&'static HarmonyEncoding> {
    static ENCODING: LazyLock<anyhow::Result<HarmonyEncoding>> = LazyLock::new(|| {
        load_harmony_encoding(HarmonyEncodingName::HarmonyGptOss)
            .context("failed to load harmony encoding for gpt-oss")
    });

-    ENCODING.as_ref().map_err(|error| Error::HarmonyOutputParsing {
+    ENCODING.as_ref().map_err(|error| Error::HarmonyEncoding {
        error: error.to_report_string().into(),
    })
 }
@@ -3,7 +3,7 @@

 //! Native Harmony chat renderer for `gpt_oss`.

-pub(crate) mod encoding;
+mod encoding;

 use openai_harmony::HarmonyEncoding;
 use openai_harmony::chat::{
@@ -12,14 +12,16 @@ use openai_harmony::chat::{
 };
 use thiserror_ext::AsReport as _;
 use time::macros::format_description;
-use vllm_text::Prompt;

-use self::encoding::harmony_encoding;
-use super::{ChatRenderer, RenderedPrompt, request_template_kwargs};
+pub use self::encoding::harmony_encoding;
+use super::{
+    ChatRenderer, RenderRequest, RenderedPrompt, RenderedPromptContent, request_template_kwargs,
+};
 use crate::error::{Error, Result};
-use crate::event::AssistantContentBlock;
-use crate::request::{ChatContent, ChatMessage, ChatRequest, ChatTool, GenerationPromptMode};
-use crate::{AssistantMessageExt as _, ReasoningEffort};
+use crate::{
+    AssistantContentBlock, AssistantMessageExt as _, ChatContent, ChatMessage,
+    GenerationPromptMode, ReasoningEffort, Tool,
+};

 const SYSTEM_START_DATE_ENV: &str = "VLLM_SYSTEM_START_DATE";
 const HARMONY_SYSTEM_INSTRUCTIONS_ENV: &str = "VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS";
@@ -73,7 +75,7 @@ impl HarmonyChatRenderer {
    ///
    /// Harmony owns both prompt formatting and tokenization, so the Rust
    /// frontend bypasses the generic HF tokenizer path for GPT-OSS input.
-    fn render_token_ids(&self, request: &ChatRequest) -> Result<Vec<u32>> {
+    fn render_token_ids(&self, request: &RenderRequest<'_>) -> Result<Vec<u32>> {
        if request.has_multimodal() {
            return Err(Error::UnsupportedMultimodalContent("image_url"));
        }
@@ -111,12 +113,12 @@ impl HarmonyChatRenderer {
 }

 impl ChatRenderer for HarmonyChatRenderer {
-    /// Render a chat request as [`Prompt::TokenIds`] with template kwargs echoed
-    /// for downstream accounting/debugging.
-    fn render(&self, request: &ChatRequest) -> Result<RenderedPrompt> {
+    /// Render a chat request as token IDs with template kwargs echoed for
+    /// downstream accounting/debugging.
+    fn render(&self, request: RenderRequest<'_>) -> Result<RenderedPrompt> {
        Ok(RenderedPrompt {
-            prompt: Prompt::TokenIds(self.render_token_ids(request)?),
-            effective_template_kwargs: request_template_kwargs(request),
+            content: RenderedPromptContent::TokenIds(self.render_token_ids(&request)?),
+            effective_template_kwargs: request_template_kwargs(&request),
        })
    }
 }
@@ -126,10 +128,10 @@ impl ChatRenderer for HarmonyChatRenderer {
 /// This adds the Harmony system/developer preamble, peels at most one leading
 /// system/developer instruction message, and then lowers the remaining chat
 /// history message-by-message.
-fn to_harmony_messages(request: &ChatRequest, options: &Options) -> Result<Vec<Message>> {
+fn to_harmony_messages(request: &RenderRequest<'_>, options: &Options) -> Result<Vec<Message>> {
    let (instructions, leading_developer_tools, remaining_messages) =
-        peel_leading_instructions(&request.messages)?;
-    let tool_call_names = tool_call_names(&request.messages);
+        peel_leading_instructions(request.messages)?;
+    let tool_call_names = tool_call_names(request.messages);
    let mut messages =
        build_harmony_preamble(request, instructions, leading_developer_tools, options)?;

@@ -147,7 +149,7 @@ fn to_harmony_messages(request: &ChatRequest, options: &Options) -> Result<Vec<M
 #[allow(clippy::type_complexity)]
 fn peel_leading_instructions(
    messages: &[ChatMessage],
-) -> Result<(Option<String>, Option<&[ChatTool]>, &[ChatMessage])> {
+) -> Result<(Option<String>, Option<&[Tool]>, &[ChatMessage])> {
    let Some(first) = messages.first() else {
        return Ok((None, None, messages));
    };
@@ -172,9 +174,9 @@ fn peel_leading_instructions(
 /// a developer message depending on `use_system_instructions`; request-level and
 /// leading developer tools are attached to the developer message.
 fn build_harmony_preamble(
-    request: &ChatRequest,
+    request: &RenderRequest<'_>,
    instructions: Option<String>,
-    leading_developer_tools: Option<&[ChatTool]>,
+    leading_developer_tools: Option<&[Tool]>,
    options: &Options,
 ) -> Result<Vec<Message>> {
    let mut messages = vec![Message::from_role_and_content(
@@ -211,12 +213,12 @@ fn build_harmony_preamble(

 /// Collect request-level and leading developer function tools for the preamble.
 fn preamble_tool_descriptions(
-    request: &ChatRequest,
-    leading_developer_tools: Option<&[ChatTool]>,
+    request: &RenderRequest<'_>,
+    leading_developer_tools: Option<&[Tool]>,
 ) -> Vec<ToolDescription> {
    let mut tools = Vec::new();
    if request.tool_parsing_enabled() {
-        tools.extend(to_tool_descriptions(&request.tools));
+        tools.extend(to_tool_descriptions(request.tools));
    }
    if let Some(leading_developer_tools) = leading_developer_tools {
        tools.extend(to_tool_descriptions(leading_developer_tools));
@@ -311,7 +313,7 @@ fn to_harmony_message(
 fn system_or_developer_message(
    role: &str,
    instructions: String,
-    tools: Option<&[ChatTool]>,
+    tools: Option<&[Tool]>,
    options: &Options,
 ) -> Result<Message> {
    if role == "system" && options.use_system_instructions {
@@ -325,7 +327,7 @@ fn system_or_developer_message(
 }

 /// Build a Harmony developer message with optional instructions and function tools.
-fn developer_message(instructions: Option<String>, tools: Option<&[ChatTool]>) -> Message {
+fn developer_message(instructions: Option<String>, tools: Option<&[Tool]>) -> Message {
    let mut content = DeveloperContent::new();
    if let Some(instructions) = instructions.filter(|text| !text.is_empty()) {
        content = content.with_instructions(instructions);
@@ -427,11 +429,11 @@ fn auto_drop_analysis_messages(messages: Vec<Message>) -> Vec<Message> {

 /// Flatten vLLM text content and reject unsupported multimodal parts.
 fn flatten_text(content: &ChatContent) -> Result<String> {
-    content.try_flatten_to_text()
+    content.try_flatten_to_text().map_err(Error::UnsupportedMultimodalContent)
 }

 /// Convert vLLM function tool definitions to Harmony tool descriptions.
-fn to_tool_descriptions(tools: &[ChatTool]) -> Vec<ToolDescription> {
+fn to_tool_descriptions(tools: &[Tool]) -> Vec<ToolDescription> {
    tools
        .iter()
        .map(|tool| {
@@ -10,15 +10,15 @@ use super::HarmonyChatRenderer;
 use super::encoding::harmony_encoding;
 use crate::ChatRenderer;
 use crate::error::Error;
-use crate::event::{AssistantContentBlock, AssistantToolCall};
-use crate::renderer::test_utils::{FixtureRequestOptions, fixture_chat_request};
-use crate::request::{
-    ChatContentPart, ChatMessage, ChatRequest, GenerationPromptMode, ReasoningEffort,
+use crate::test_utils::{FixtureRequestOptions, fixture_chat_request};
+use crate::{AssistantContentBlock, AssistantToolCall};
+use crate::{
+    ChatContentPart, ChatMessage, GenerationPromptMode, ReasoningEffort, TestRenderRequest,
 };

 const PINNED_DATE: &str = "2025-06-28";

-fn fixture_request(input_name: &str) -> ChatRequest {
+fn fixture_request(input_name: &str) -> TestRenderRequest {
    fixture_chat_request(
        &fixture_path(input_name),
        FixtureRequestOptions {
@@ -30,7 +30,7 @@ fn fixture_request(input_name: &str) -> ChatRequest {

 fn fixture_path(name: &str) -> PathBuf {
    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
-        .join("src/renderer/harmony")
+        .join("src/harmony")
        .join("fixtures")
        .join(name)
 }
@@ -39,24 +39,24 @@ fn test_renderer(use_system_instructions: bool) -> HarmonyChatRenderer {
    HarmonyChatRenderer::with_options(PINNED_DATE, use_system_instructions).unwrap()
 }

-fn render_token_ids(request: &ChatRequest) -> Vec<u32> {
+fn render_token_ids(request: &TestRenderRequest) -> Vec<u32> {
    render_token_ids_with(&test_renderer(false), request)
 }

-fn render_token_ids_with(renderer: &HarmonyChatRenderer, request: &ChatRequest) -> Vec<u32> {
+fn render_token_ids_with(renderer: &HarmonyChatRenderer, request: &TestRenderRequest) -> Vec<u32> {
    renderer
-        .render(request)
+        .render(request.as_request())
        .unwrap()
-        .prompt
+        .content
        .into_token_ids()
        .expect("Harmony renderer returns token IDs")
 }

-fn render_prompt_text(request: &ChatRequest) -> String {
+fn render_prompt_text(request: &TestRenderRequest) -> String {
    render_prompt_text_with(&test_renderer(false), request)
 }

-fn render_prompt_text_with(renderer: &HarmonyChatRenderer, request: &ChatRequest) -> String {
+fn render_prompt_text_with(renderer: &HarmonyChatRenderer, request: &TestRenderRequest) -> String {
    let token_ids = render_token_ids_with(renderer, request);
    harmony_encoding().unwrap().tokenizer().decode_utf8(&token_ids).unwrap()
 }
@@ -137,10 +137,10 @@ fn drops_stale_analysis_fixture() {

 #[test]
 fn rejects_invalid_reasoning_effort() {
-    let mut request = ChatRequest::for_test();
+    let mut request = TestRenderRequest::for_test();
    request.chat_options.reasoning_effort = Some(ReasoningEffort::None);

-    let error = test_renderer(false).render(&request).unwrap_err();
+    let error = test_renderer(false).render(request.as_request()).unwrap_err();

    expect![[r#"chat template error: reasoning_effort="none" is not supported by Harmony. Supported values are: low, medium, high."#]]
        .assert_eq(&error.to_report_string());
@@ -148,7 +148,7 @@ fn rejects_invalid_reasoning_effort() {

 #[test]
 fn rejects_unknown_tool_response_id() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![
            ChatMessage::assistant_blocks(vec![AssistantContentBlock::ToolCall(
                AssistantToolCall {
@@ -159,10 +159,10 @@ fn rejects_unknown_tool_response_id() {
            )]),
            ChatMessage::tool_response("{}", "call-unknown"),
        ],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

-    let error = test_renderer(false).render(&request).unwrap_err();
+    let error = test_renderer(false).render(request.as_request()).unwrap_err();

    expect![
        "chat template error: invalid Harmony tool message: unknown tool_call_id `call-unknown`"
@@ -172,14 +172,14 @@ fn rejects_unknown_tool_response_id() {

 #[test]
 fn rejects_multimodal_input() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![ChatMessage::user(vec![ChatContentPart::image_url(
            "data:image/png;base64,test",
        )])],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

-    let error = test_renderer(false).render(&request).unwrap_err();
+    let error = test_renderer(false).render(request.as_request()).unwrap_err();

    assert!(matches!(
        error,
@@ -189,16 +189,16 @@ fn rejects_multimodal_input() {

 #[test]
 fn rejects_continue_final_assistant() {
-    let mut request = ChatRequest {
+    let mut request = TestRenderRequest {
        messages: vec![
            ChatMessage::user("write"),
            ChatMessage::assistant_text("partial"),
        ],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };
    request.chat_options.generation_prompt_mode = GenerationPromptMode::ContinueFinalAssistant;

-    let error = test_renderer(false).render(&request).unwrap_err();
+    let error = test_renderer(false).render(request.as_request()).unwrap_err();

    expect!["chat template error: Harmony renderer does not support continue_final_message"]
        .assert_eq(&error.to_report_string());
@@ -33,8 +33,11 @@ pub enum ChatTemplateContentFormatOption {
 }

 impl ChatTemplateContentFormatOption {
+    /// CLI/config literal for automatic content-format detection.
    pub const AUTO_LITERAL: &str = "auto";
+    /// CLI/config literal for OpenAI-compatible structured content.
    pub const OPENAI_LITERAL: &str = "openai";
+    /// CLI/config literal for flattened string content.
    pub const STRING_LITERAL: &str = "string";
 }

@@ -1,27 +1,31 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

+//! Hugging Face Jinja chat-template renderer.
+
 use std::collections::HashMap;
+use std::path::Path;

 use serde::Serialize;
 use serde_json::Value as JsonValue;
 use thiserror_ext::AsReport as _;
 use tracing::{info, trace, warn};
-use vllm_text::Prompt;
-use vllm_text::backend::hf::{
-    HfSpecialTokens, HfTokenizerConfig, ResolvedModelFiles, load_tokenizer_config,
-};
+use vllm_tokenizer::{HfSpecialTokens, HfTokenizerConfig, load_tokenizer_config};

 use self::format::{
    ChatTemplateContentFormat, ChatTemplateContentFormatOption as ContentFormatOption,
 };
-use self::template::{CompiledChatTemplate, TemplateContext};
+use self::template::{
+    CompiledChatTemplate, TemplateContext, load_chat_template, resolve_chat_template,
+};
 use self::value::{TemplateValue, to_template_value};
-use super::{ChatRenderer, RenderedPrompt, effective_template_kwargs};
+use super::{
+    ChatRenderer, RenderRequest, RenderedPrompt, RenderedPromptContent, effective_template_kwargs,
+};
 use crate::error::Result;
-use crate::request::{ChatContent, ChatContentPart, ChatMessage, ChatRequest};
 use crate::{
-    AssistantContentBlock, AssistantMessageExt, ChatTool, Error, LoadModelBackendsOptions,
+    AssistantContentBlock, AssistantMessageExt, ChatContent, ChatContentPart, ChatMessage, Error,
+    Tool,
 };

 mod error;
@@ -30,8 +34,6 @@ mod template;
 mod tojson;
 mod value;

-pub use template::{load_chat_template, resolve_chat_template};
-
 pub use self::format::ChatTemplateContentFormatOption;

 /// Template-visible placeholder tokens per supported modality.
@@ -40,11 +42,36 @@ pub use self::format::ChatTemplateContentFormatOption;
 /// content parts of that modality are rejected during rendering.
 #[derive(Debug, Clone, Default)]
 pub struct MultimodalRenderInfo {
+    /// Template-visible image placeholder token.
    pub image_token: Option<String>,
+    /// Template-visible video placeholder token.
    pub video_token: Option<String>,
+    /// Template-visible audio placeholder token.
    pub audio_token: Option<String>,
 }

+/// Resolved local files consumed while constructing an HF renderer.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct HfRendererFiles<'a> {
+    /// Resolved `tokenizer_config.json`, when available.
+    pub tokenizer_config: Option<&'a Path>,
+    /// Resolved dedicated chat-template file, when available.
+    pub chat_template: Option<&'a Path>,
+}
+
+/// Model-level options used to construct an HF renderer.
+#[derive(Debug, Clone, Default)]
+pub struct HfRendererConfig {
+    /// Optional inline or file-backed chat-template override.
+    pub chat_template: Option<String>,
+    /// Default template kwargs merged before request-level kwargs.
+    pub default_template_kwargs: HashMap<String, JsonValue>,
+    /// How `message.content` should be represented in the template.
+    pub content_format: ContentFormatOption,
+    /// Template-visible multimodal placeholder tokens.
+    pub multimodal: Option<MultimodalRenderInfo>,
+}
+
 /// Hugging Face chat-template renderer backed by the local Jinja chat-template
 /// state.
 pub struct HfChatRenderer {
@@ -76,27 +103,25 @@ impl HfChatRenderer {
        })
    }

+    /// Attach named special tokens exposed to chat templates.
    pub fn with_special_tokens(mut self, special_tokens: Option<HfSpecialTokens>) -> Self {
        self.special_tokens = special_tokens;
        self
    }

+    /// Attach template-visible multimodal placeholder tokens.
    pub fn with_multimodal(mut self, multimodal: Option<MultimodalRenderInfo>) -> Self {
        self.multimodal = multimodal;
        self
    }

-    /// Create a renderer from the given model files and loading options.
-    pub fn load(
-        files: &ResolvedModelFiles,
-        options: LoadModelBackendsOptions,
-        multimodal: Option<MultimodalRenderInfo>,
-    ) -> Result<Self> {
+    /// Create a renderer from resolved local model files and renderer options.
+    pub fn load(files: HfRendererFiles<'_>, options: HfRendererConfig) -> Result<Self> {
        let HfTokenizerConfig {
            special_tokens,
            chat_template,
            ..
-        } = load_tokenizer_config(files.tokenizer_config_path.as_deref())?;
+        } = load_tokenizer_config(files.tokenizer_config)?;
        let mut template = chat_template;
        let special_tokens = (!special_tokens.is_empty()).then_some(special_tokens);

@@ -106,7 +131,7 @@ impl HfChatRenderer {
                    .map_err(|error| Error::ChatTemplate(error.to_report_string()))?,
            );
            info!("using configured chat template override");
-        } else if let Some(chat_template_path) = files.chat_template_path.as_deref() {
+        } else if let Some(chat_template_path) = files.chat_template {
            // If independent chat template file(s) exist and contain non-empty content,
            // they take priority over template entries in the tokenizer config
            let file_template = load_chat_template(chat_template_path)
@@ -128,11 +153,11 @@ impl HfChatRenderer {

        Ok(Self::new(
            template,
-            options.default_chat_template_kwargs,
-            options.chat_template_content_format,
+            options.default_template_kwargs,
+            options.content_format,
        )?
        .with_special_tokens(special_tokens)
-        .with_multimodal(multimodal))
+        .with_multimodal(options.multimodal))
    }

    /// Apply the chat template to one chat request, rendering the prompt string
@@ -141,7 +166,7 @@ impl HfChatRenderer {
    /// If the request carries a per-request `chat_template` override, a
    /// temporary template is compiled from that string and used instead of
    /// the model's default.
-    fn apply_chat_template(&self, request: &ChatRequest) -> Result<RenderedPrompt> {
+    fn apply_chat_template(&self, request: &RenderRequest<'_>) -> Result<RenderedPrompt> {
        let override_template = request
            .chat_options
            .chat_template
@@ -162,10 +187,10 @@ impl HfChatRenderer {
    fn apply_chat_template_inner(
        &self,
        effective_template: &CompiledChatTemplate,
-        request: &ChatRequest,
+        request: &RenderRequest<'_>,
    ) -> Result<RenderedPrompt> {
        let mut messages = to_template_messages(
-            &request.messages,
+            request.messages,
            effective_template.content_format(),
            self.multimodal.as_ref(),
        )?;
@@ -181,7 +206,7 @@ impl HfChatRenderer {
            None
        };

-        let tools = request.tool_parsing_enabled().then(|| to_template_tools(&request.tools));
+        let tools = request.tool_parsing_enabled().then(|| to_template_tools(request.tools));
        trace!(
            message_count = messages.len(),
            content_format = ?effective_template.content_format(),
@@ -198,7 +223,7 @@ impl HfChatRenderer {
                add_generation_prompt: request.chat_options.add_generation_prompt(),
                continue_final_message: request.chat_options.continue_final_message(),
                tools: tools.as_deref(),
-                documents: request.documents.as_deref(),
+                documents: request.documents,
                template_kwargs: Some(&effective_template_kwargs),
                special_tokens: self.special_tokens.as_ref(),
            })
@@ -217,15 +242,15 @@ impl HfChatRenderer {
        );

        Ok(RenderedPrompt {
-            prompt: Prompt::Text(prompt),
+            content: RenderedPromptContent::Text(prompt),
            effective_template_kwargs,
        })
    }
 }

 impl ChatRenderer for HfChatRenderer {
-    fn render(&self, request: &ChatRequest) -> Result<RenderedPrompt> {
-        self.apply_chat_template(request)
+    fn render(&self, request: RenderRequest<'_>) -> Result<RenderedPrompt> {
+        self.apply_chat_template(&request)
    }
 }

@@ -556,7 +581,7 @@ fn truncate_prompt_at_continue_final_message_tag(
    Ok(rendered)
 }

-fn to_template_tools(tools: &[ChatTool]) -> Vec<TemplateTool> {
+fn to_template_tools(tools: &[Tool]) -> Vec<TemplateTool> {
    tools
        .iter()
        .map(|tool| TemplateTool {
@@ -577,42 +602,39 @@ mod tests {

    use expect_test::expect;
    use serde_json::Value;
-    use vllm_text::Prompt;
-    use vllm_text::backend::hf::{HfSpecialTokens, NamedSpecialToken};
+    use vllm_tokenizer::{HfSpecialTokens, NamedSpecialToken};

    use super::{ChatTemplateContentFormatOption, HfChatRenderer, MultimodalRenderInfo};
-    use crate::request::{
-        ChatContentPart, ChatMessage, ChatRequest, ChatRole, ChatTool, ChatToolChoice,
-        GenerationPromptMode, ReasoningEffort,
+    use crate::{
+        AssistantContentBlock, ChatContentPart, ChatMessage, ChatRenderer, ChatRole,
+        ChatToolChoice, Error, GenerationPromptMode, ReasoningEffort, RenderedPromptContent,
+        Result, TestRenderRequest, Tool,
    };
-    use crate::{AssistantContentBlock, ChatRenderer, Error, Result};

-    const QWEN3_0_6B_TEMPLATE: &str = include_str!("../../../tests/templates/qwen3.jinja");
-    const QWEN3_5_0_8B_TEMPLATE: &str = include_str!("../../../tests/templates/qwen35.jinja");
+    const QWEN3_0_6B_TEMPLATE: &str = include_str!("../../../chat/tests/templates/qwen3.jinja");
+    const QWEN3_5_0_8B_TEMPLATE: &str = include_str!("../../../chat/tests/templates/qwen35.jinja");

-    fn sample_request(messages: Vec<ChatMessage>) -> ChatRequest {
-        ChatRequest {
+    fn sample_request(messages: Vec<ChatMessage>) -> TestRenderRequest {
+        TestRenderRequest {
            messages,
-            request_id: "render-test".to_string(),
-            ..ChatRequest::for_test()
+            ..TestRenderRequest::for_test()
        }
    }

-    fn render(template: Option<&str>, request: &ChatRequest) -> Result<String> {
-        HfChatRenderer::new(
+    fn render(template: Option<&str>, request: &TestRenderRequest) -> Result<String> {
+        let rendered = HfChatRenderer::new(
            template.map(str::to_owned),
            HashMap::new(),
            ChatTemplateContentFormatOption::Auto,
        )?
-        .render(request)?
-        .prompt
-        .into_text()
-        .map_err(|_| unreachable!("HF renderer should return text prompt"))
+        .render(request.as_request())?
+        .content;
+        Ok(rendered.into_text().expect("HF renderer should return text prompt"))
    }

    fn render_mm(
        template: &str,
-        request: &ChatRequest,
+        request: &TestRenderRequest,
        content_format: ChatTemplateContentFormatOption,
    ) -> Result<crate::RenderedPrompt> {
        HfChatRenderer::new(Some(template.to_string()), HashMap::new(), content_format)?
@@ -621,10 +643,10 @@ mod tests {
                video_token: Some("<video>".to_string()),
                audio_token: Some("<audio>".to_string()),
            }))
-            .render(request)
+            .render(request.as_request())
    }

-    fn image_request() -> ChatRequest {
+    fn image_request() -> TestRenderRequest {
        sample_request(vec![ChatMessage::user(vec![
            ChatContentPart::text("a"),
            ChatContentPart::image_url("data:image/png;base64,test"),
@@ -632,7 +654,7 @@ mod tests {
        ])])
    }

-    fn video_request() -> ChatRequest {
+    fn video_request() -> TestRenderRequest {
        sample_request(vec![ChatMessage::user(vec![
            ChatContentPart::text("a"),
            ChatContentPart::video_url("https://example.com/demo.mp4"),
@@ -640,7 +662,7 @@ mod tests {
        ])])
    }

-    fn audio_request() -> ChatRequest {
+    fn audio_request() -> TestRenderRequest {
        sample_request(vec![ChatMessage::user(vec![
            ChatContentPart::text("a"),
            ChatContentPart::input_audio("dGVzdA==", Some("wav".to_string())),
@@ -658,7 +680,10 @@ mod tests {
        )
        .unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("a<image>b".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("a<image>b".to_string())
+        );
    }

    #[test]
@@ -670,7 +695,10 @@ mod tests {
        )
        .unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("a<video>b".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("a<video>b".to_string())
+        );
    }

    #[test]
@@ -683,8 +711,8 @@ mod tests {
        .unwrap();

        assert_eq!(
-            rendered.prompt,
-            Prompt::Text("a<audio><audio>b".to_string())
+            rendered.content,
+            RenderedPromptContent::Text("a<audio><audio>b".to_string())
        );
    }

@@ -697,7 +725,10 @@ mod tests {
        )
        .unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("a<|image_pad|>b".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("a<|image_pad|>b".to_string())
+        );
    }

    #[test]
@@ -709,7 +740,10 @@ mod tests {
        )
        .unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("a<|video_pad|>b".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("a<|video_pad|>b".to_string())
+        );
    }

    #[test]
@@ -722,8 +756,8 @@ mod tests {
        .unwrap();

        assert_eq!(
-            rendered.prompt,
-            Prompt::Text("a<|audio_pad|><|audio_pad|>b".to_string())
+            rendered.content,
+            RenderedPromptContent::Text("a<|audio_pad|><|audio_pad|>b".to_string())
        );
    }

@@ -740,7 +774,7 @@ mod tests {
            video_token: None,
            audio_token: None,
        }))
-        .render(&video_request())
+        .render(video_request().as_request())
        .unwrap_err();

        assert!(matches!(
@@ -842,12 +876,12 @@ mod tests {
            ChatTemplateContentFormatOption::OpenAi,
        )
        .unwrap()
-        .prompt;
+        .content;

        // Anything rendered after the continued text (here the image
        // placeholder and the end marker) is truncated away, matching
        // transformers.
-        assert_eq!(rendered, Prompt::Text("Sure,".to_string()));
+        assert_eq!(rendered, RenderedPromptContent::Text("Sure,".to_string()));
    }

    #[test]
@@ -904,7 +938,7 @@ mod tests {
    fn chat_template_exposes_developer_tools() {
        let request = sample_request(vec![ChatMessage::developer(
            "policy",
-            Some(vec![ChatTool {
+            Some(vec![Tool {
                name: "get_weather".to_string(),
                description: Some("Get weather".to_string()),
                parameters: serde_json::json!({
@@ -1005,10 +1039,13 @@ mod tests {
        )
        .unwrap()
        .with_special_tokens(Some(special_tokens))
-        .apply_chat_template(&request)
+        .apply_chat_template(&request.as_request())
        .unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("<bos>|true".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("<bos>|true".to_string())
+        );
    }

    #[test]
@@ -1046,11 +1083,14 @@ mod tests {
            ChatTemplateContentFormatOption::String,
        )
        .unwrap()
-        .render(&request)
+        .render(request.as_request())
        .unwrap()
-        .prompt;
+        .content;

-        assert_eq!(rendered, Prompt::Text("hello world".to_string()));
+        assert_eq!(
+            rendered,
+            RenderedPromptContent::Text("hello world".to_string())
+        );
    }

    #[test]
@@ -1066,11 +1106,14 @@ mod tests {
            ChatTemplateContentFormatOption::OpenAi,
        )
        .unwrap()
-        .render(&request)
+        .render(request.as_request())
        .unwrap()
-        .prompt;
+        .content;

-        assert_eq!(rendered, Prompt::Text("hello world".to_string()));
+        assert_eq!(
+            rendered,
+            RenderedPromptContent::Text("hello world".to_string())
+        );
    }

    #[test]
@@ -1091,9 +1134,9 @@ mod tests {
        )
        .unwrap();

-        let rendered = renderer.render(&request).unwrap().prompt;
+        let rendered = renderer.render(request.as_request()).unwrap().content;

-        assert_eq!(rendered, Prompt::Text("true|x".to_string()));
+        assert_eq!(rendered, RenderedPromptContent::Text("true|x".to_string()));
    }

    #[test]
@@ -1115,9 +1158,12 @@ mod tests {
        )
        .unwrap();

-        let rendered = renderer.render(&request).unwrap();
+        let rendered = renderer.render(request.as_request()).unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("max".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("max".to_string())
+        );
        assert_eq!(
            rendered.effective_template_kwargs.get("reasoning_effort"),
            Some(&Value::String("max".to_string()))
@@ -1144,9 +1190,12 @@ mod tests {
        )
        .unwrap();

-        let rendered = renderer.render(&request).unwrap();
+        let rendered = renderer.render(request.as_request()).unwrap();

-        assert_eq!(rendered.prompt, Prompt::Text("none|true".to_string()));
+        assert_eq!(
+            rendered.content,
+            RenderedPromptContent::Text("none|true".to_string())
+        );
        assert_eq!(
            rendered.effective_template_kwargs.get("reasoning_effort"),
            Some(&Value::String("none".to_string()))
@@ -1222,7 +1271,7 @@ mod tests {
    #[test]
    fn chat_template_exposes_tools_to_templates_when_auto_enabled() {
        let mut request = sample_request(vec![ChatMessage::text(ChatRole::User, "hello")]);
-        request.tools = vec![ChatTool {
+        request.tools = vec![Tool {
            name: "get_weather".to_string(),
            description: Some("Get weather".to_string()),
            parameters: serde_json::json!({
@@ -1326,7 +1375,7 @@ mod tests {
                "<|im_start|>user\na<|vision_start|><|image_pad|><|vision_end|>b<|im_end|>\n",
            )
        "#]]
-        .assert_debug_eq(&rendered.prompt);
+        .assert_debug_eq(&rendered.content);
    }

    #[test]
@@ -5,7 +5,7 @@
 //!
 //! This module is inlined from SMG's tokenizer crate with local adaptations:
 //! - thinking-related detection/state is removed
-//! - special tokens are wired to `vllm_text::backends::hf::HfSpecialTokens`
+//! - special tokens are wired to `vllm_tokenizer::HfSpecialTokens`

 use std::collections::HashMap;
 use std::fs;
@@ -14,14 +14,14 @@ use std::path::Path;
 use minijinja::Environment;
 use serde::{Deserialize, Serialize};
 use serde_json::{self};
-use vllm_text::backend::hf::HfSpecialTokens;
+use vllm_tokenizer::HfSpecialTokens;

 use super::error::TemplateError;
 use super::format::{
    ChatTemplateContentFormat, ChatTemplateContentFormatOption, detect_chat_template_content_format,
 };
 use super::tojson::hf_tojson_filter;
-use crate::renderer::hf::{TemplateMessage, TemplateTool};
+use crate::hf::{TemplateMessage, TemplateTool};

 type Result<T> = std::result::Result<T, TemplateError>;

@@ -55,7 +55,7 @@ pub(super) struct TemplateContext<'a> {
 }

 /// Load chat template from a file (`.jinja` or `.json` containing Jinja).
-pub fn load_chat_template(template_path: &Path) -> Result<Option<String>> {
+pub(crate) fn load_chat_template(template_path: &Path) -> Result<Option<String>> {
    let content = fs::read_to_string(template_path).map_err(TemplateError::ReadTemplateFile)?;

    if template_path.extension().is_some_and(|ext| ext == "json") {
@@ -82,7 +82,7 @@ pub fn load_chat_template(template_path: &Path) -> Result<Option<String>> {
 }

 /// Resolve a configured chat template value into a template string.
-pub fn resolve_chat_template(chat_template: &str) -> Result<String> {
+pub(crate) fn resolve_chat_template(chat_template: &str) -> Result<String> {
    let path = Path::new(chat_template);
    if path.exists() {
        return load_chat_template(path).map(|template| template.unwrap_or_default());
@@ -136,7 +136,7 @@ mod tests {
    use std::fs;

    use tempfile::TempDir;
-    use vllm_text::backend::hf::{HfSpecialTokens, NamedSpecialToken};
+    use vllm_tokenizer::{HfSpecialTokens, NamedSpecialToken};

    use super::*;

@@ -5,13 +5,15 @@ use std::collections::HashMap;

 use serde_json::{Map, Value, json};
 use thiserror_ext::AsReport as _;
-use vllm_text::Prompt;
-use vllm_text::tokenizer::{DynTokenizer, Tokenizer};
+use vllm_tokenizer::{DynTokenizer, Tokenizer};

-use super::{ChatRenderer, RenderedPrompt, request_template_kwargs};
+use super::{
+    ChatRenderer, RenderRequest, RenderedPrompt, RenderedPromptContent, request_template_kwargs,
+};
 use crate::error::{Error, Result};
-use crate::request::{ChatContent, ChatContentPart, ChatMessage, ChatRequest, ChatTool};
-use crate::{AssistantContentBlock, AssistantToolCall};
+use crate::{
+    AssistantContentBlock, AssistantToolCall, ChatContent, ChatContentPart, ChatMessage, Tool,
+};

 const MESSAGE_USER: &str = "<|message_user|>";
 const MESSAGE_MODEL: &str = "<|message_model|>";
@@ -170,7 +172,7 @@ impl InklingChatRenderer {
        )
    }

-    fn write_tool_declarations(&self, out: &mut Vec<u32>, tools: &[&ChatTool]) -> Result<()> {
+    fn write_tool_declarations(&self, out: &mut Vec<u32>, tools: &[&Tool]) -> Result<()> {
        if tools.is_empty() {
            return Ok(());
        }
@@ -277,14 +279,14 @@ impl InklingChatRenderer {
        tool_call_id: &str,
        tool_call_id_to_name: &HashMap<String, String>,
    ) -> Result<()> {
-        let text = content.try_flatten_to_text()?;
+        let text = content.try_flatten_to_text().map_err(Error::UnsupportedMultimodalContent)?;
        let tool_name = tool_call_id_to_name.get(tool_call_id).map(String::as_str).unwrap_or("");
        self.write_text_block(out, self.special.message_tool, Some(tool_name), &text)
    }
 }

 impl ChatRenderer for InklingChatRenderer {
-    fn render(&self, request: &ChatRequest) -> Result<RenderedPrompt> {
+    fn render(&self, request: RenderRequest<'_>) -> Result<RenderedPrompt> {
        request.validate()?;
        if request.chat_options.continue_final_message() {
            return Err(Error::ChatTemplate(
@@ -294,13 +296,13 @@ impl ChatRenderer for InklingChatRenderer {

        let mut out = Vec::new();
        let mut tool_call_id_to_name = HashMap::new();
-        let effective_template_kwargs = request_template_kwargs(request);
-        let tools = rendered_tools(request);
+        let effective_template_kwargs = request_template_kwargs(&request);
+        let tools = rendered_tools(&request);
        self.write_tool_declarations(&mut out, &tools)?;
        let mut reasoning_effort =
            resolve_reasoning_effort(effective_template_kwargs.get("reasoning_effort"));

-        for message in &request.messages {
+        for message in request.messages {
            if !matches!(
                message,
                ChatMessage::System { .. } | ChatMessage::Developer { .. }
@@ -345,7 +347,7 @@ impl ChatRenderer for InklingChatRenderer {
        }

        Ok(RenderedPrompt {
-            prompt: Prompt::TokenIds(out),
+            content: RenderedPromptContent::TokenIds(out),
            effective_template_kwargs,
        })
    }
@@ -378,14 +380,14 @@ fn resolve_special_token(tokenizer: &dyn Tokenizer, token: &str) -> Result<u32>
    })
 }

-fn rendered_tools(request: &ChatRequest) -> Vec<&ChatTool> {
+fn rendered_tools<'a>(request: &RenderRequest<'a>) -> Vec<&'a Tool> {
    if !request.tool_parsing_enabled() {
        return Vec::new();
    }

    let mut tools = Vec::with_capacity(request.tools.len());
    tools.extend(request.tools.iter());
-    for message in &request.messages {
+    for message in request.messages {
        if let ChatMessage::Developer {
            tools: Some(local_tools),
            ..
@@ -7,27 +7,23 @@ use std::sync::Arc;
 use expect_test::{ExpectFile, expect_file};
 use serde_json::json;
 use thiserror_ext::AsReport;
-use vllm_text::tokenizer::Tokenizer;
+use vllm_tokenizer::Tokenizer;

-use crate::renderer::test_utils::{FixtureRequestOptions, fixture_chat_request};
+use crate::test_utils::{FixtureRequestOptions, fixture_chat_request};

 use super::{
    AUDIO_END, CONTENT_AUDIO_INPUT, CONTENT_IMAGE, CONTENT_INVOKE_TOOL_JSON, CONTENT_TEXT,
    CONTENT_THINKING, CONTENT_XML, END_MESSAGE, InklingChatRenderer, MESSAGE_MODEL, MESSAGE_SYSTEM,
    MESSAGE_TOOL, MESSAGE_USER,
 };
-use crate::event::{AssistantContentBlock, AssistantToolCall};
-use crate::request::{ChatMessage, ChatRequest, GenerationPromptMode, ReasoningEffort};
+use crate::{AssistantContentBlock, AssistantToolCall};
+use crate::{ChatMessage, GenerationPromptMode, ReasoningEffort, TestRenderRequest};
 use crate::{ChatRenderer, Error};

 struct FixtureTokenizer;

 impl Tokenizer for FixtureTokenizer {
-    fn encode(
-        &self,
-        text: &str,
-        _add_special_tokens: bool,
-    ) -> vllm_text::tokenizer::Result<Vec<u32>> {
+    fn encode(&self, text: &str, _add_special_tokens: bool) -> vllm_tokenizer::Result<Vec<u32>> {
        Ok(text.bytes().map(u32::from).collect())
    }

@@ -35,7 +31,7 @@ impl Tokenizer for FixtureTokenizer {
        &self,
        token_ids: &[u32],
        _skip_special_tokens: bool,
-    ) -> vllm_text::tokenizer::Result<String> {
+    ) -> vllm_tokenizer::Result<String> {
        Ok(token_ids
            .iter()
            .map(|token_id| char::from_u32(*token_id).unwrap_or('\u{FFFD}'))
@@ -86,16 +82,16 @@ fn renderer() -> InklingChatRenderer {
    InklingChatRenderer::new(Arc::new(FixtureTokenizer)).unwrap()
 }

-fn render_token_ids(request: &ChatRequest) -> Vec<u32> {
+fn render_token_ids(request: &TestRenderRequest) -> Vec<u32> {
    renderer()
-        .render(request)
+        .render(request.as_request())
        .unwrap()
-        .prompt
+        .content
        .into_token_ids()
        .expect("Inkling renderer returns token IDs")
 }

-fn fixture_request(name: &str) -> ChatRequest {
+fn fixture_request(name: &str) -> TestRenderRequest {
    fixture_chat_request(&fixture_path(name), inkling_fixture_options())
 }

@@ -108,7 +104,7 @@ fn inkling_fixture_options() -> FixtureRequestOptions {

 fn fixture_path(name: &str) -> PathBuf {
    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
-        .join("src/renderer/inkling")
+        .join("src/inkling")
        .join("fixtures")
        .join(name)
 }
@@ -229,14 +225,14 @@ fn renders_named_reasoning_effort_after_tool_declarations() {

 #[test]
 fn emits_one_reasoning_effort_for_multi_turn_conversation() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![
            ChatMessage::system("rules"),
            ChatMessage::user("user1"),
            ChatMessage::assistant_text("assistant1"),
            ChatMessage::user("user2"),
        ],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

    let rendered = render_symbolic_tokens(&render_token_ids(&request));
@@ -253,7 +249,7 @@ fn emits_one_reasoning_effort_for_multi_turn_conversation() {
 #[test]
 fn canonicalizes_numeric_zero_reasoning_effort() {
    for value in [0.0, -0.0] {
-        let mut request = ChatRequest::for_test();
+        let mut request = TestRenderRequest::for_test();
        request
            .chat_options
            .template_kwargs
@@ -270,7 +266,7 @@ fn canonicalizes_numeric_zero_reasoning_effort() {

 #[test]
 fn defaults_reasoning_effort_to_high() {
-    let request = ChatRequest::for_test();
+    let request = TestRenderRequest::for_test();

    assert!(
        render_symbolic_tokens(&render_token_ids(&request)).starts_with(
@@ -282,7 +278,7 @@ fn defaults_reasoning_effort_to_high() {

 #[test]
 fn renders_numeric_reasoning_effort_template_kwarg() {
-    let mut request = ChatRequest::for_test();
+    let mut request = TestRenderRequest::for_test();
    request
        .chat_options
        .template_kwargs
@@ -299,7 +295,7 @@ fn renders_numeric_reasoning_effort_template_kwarg() {
 #[test]
 fn ignores_unsupported_reasoning_effort_values() {
    for value in [json!(true), json!("invalid"), json!(null)] {
-        let mut request = ChatRequest::for_test();
+        let mut request = TestRenderRequest::for_test();
        request
            .chat_options
            .template_kwargs
@@ -315,38 +311,38 @@ fn ignores_unsupported_reasoning_effort_values() {
 #[test]
 fn rejects_out_of_range_reasoning_effort() {
    for value in [0.990_000_1, 1.0, 1.5, -0.1] {
-        let mut request = ChatRequest::for_test();
+        let mut request = TestRenderRequest::for_test();
        request
            .chat_options
            .template_kwargs
            .insert("reasoning_effort".to_string(), json!(value));

-        let error = renderer().render(&request).unwrap_err();
+        let error = renderer().render(request.as_request()).unwrap_err();
        assert!(error.as_report().to_string().contains("must be in [0.0, 0.99]"));
    }
 }

 #[test]
 fn rejects_continue_final_message() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![ChatMessage::assistant_text("partial")],
        chat_options: crate::ChatOptions {
            generation_prompt_mode: GenerationPromptMode::ContinueFinalAssistant,
            ..Default::default()
        },
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

-    let error = renderer().render(&request).unwrap_err();
+    let error = renderer().render(request.as_request()).unwrap_err();
    assert!(matches!(error, Error::ChatTemplate(_)));
    assert!(error.as_report().to_string().contains("continue_final_message"));
 }

 #[test]
 fn renders_developer_messages_as_system() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![ChatMessage::developer("rules", None)],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

    assert_eq!(
@@ -359,7 +355,7 @@ fn renders_developer_messages_as_system() {

 #[test]
 fn rejects_non_object_tool_call_arguments() {
-    let request = ChatRequest {
+    let request = TestRenderRequest {
        messages: vec![ChatMessage::assistant_blocks(vec![
            AssistantContentBlock::ToolCall(AssistantToolCall {
                id: "call_1".to_string(),
@@ -367,9 +363,9 @@ fn rejects_non_object_tool_call_arguments() {
                arguments: "[]".to_string(),
            }),
        ])],
-        ..ChatRequest::for_test()
+        ..TestRenderRequest::for_test()
    };

-    let error = renderer().render(&request).unwrap_err();
+    let error = renderer().render(request.as_request()).unwrap_err();
    assert!(error.as_report().to_string().contains("JSON object"));
 }
@@ -1,45 +1,70 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

+//! Engine-independent chat prompt renderers.
+//!
+//! The crate accepts borrowed chat-domain inputs and produces text or token-ID
+//! prompt artifacts. Serving requests, sampling policy, multimodal
+//! preprocessing, parser state, and engine transport remain in consumer
+//! crates.
+
 use std::collections::HashMap;
 use std::sync::Arc;

+use enum_as_inner::EnumAsInner;
+use serde::{Deserialize, Serialize};
 use serde_json::{Value, json};
-use vllm_text::Prompt;

-use crate::error::Result;
-use crate::request::{ChatRequest, ReasoningEffort};
-
-pub mod deepseek_v32;
-pub mod deepseek_v4;
+mod deepseek_v32;
+mod deepseek_v4;
+mod error;
 pub mod harmony;
 pub mod hf;
 mod inkling;
+mod request;
 mod selection;
 #[cfg(test)]
 mod test_utils;
+#[cfg(test)]
+pub(crate) use test_utils::TestRenderRequest;

 pub use deepseek_v4::DeepSeekV4ChatRenderer;
 pub use deepseek_v32::DeepSeekV32ChatRenderer;
+pub use error::{Error, Result};
 pub use harmony::HarmonyChatRenderer;
 pub use inkling::InklingChatRenderer;
+pub use request::RenderRequest;
 pub use selection::RendererSelection;
+pub use vllm_chat_types::{
+    AssistantBlockKind, AssistantContentBlock, AssistantMessage, AssistantMessageExt,
+    AssistantToolCall, ChatContent, ChatContentPart, ChatMessage, ChatOptions, ChatRole,
+    ChatToolChoice, GenerationPromptMode, ImageDetail, ReasoningEffort, Tool,
+};

-/// Rendered chat prompt submitted to the text backend.
-#[derive(Debug, Clone, PartialEq)]
+/// Engine-independent prompt content produced by a chat renderer.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, EnumAsInner)]
+#[serde(untagged)]
+pub enum RenderedPromptContent {
+    /// Untokenized prompt text.
+    Text(String),
+    /// Prompt token IDs produced by a format-specific renderer.
+    TokenIds(Vec<u32>),
+}
+
+/// Rendered chat prompt plus effective template metadata.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub struct RenderedPrompt {
-    /// The rendered prompt, either as text or already tokenized.
-    pub prompt: Prompt,
+    /// Engine-independent prompt content.
+    pub content: RenderedPromptContent,
    /// Effective chat-template kwargs visible to the renderer after applying
    /// server defaults, request overrides, and typed reasoning controls.
    pub effective_template_kwargs: HashMap<String, Value>,
 }

-/// Minimal chat-prompt renderer used by `vllm-chat`.
+/// Synchronous chat-prompt renderer.
 pub trait ChatRenderer: Send + Sync {
-    /// Render one chat request into the text prompt submitted to the text
-    /// backend.
-    fn render(&self, request: &ChatRequest) -> Result<RenderedPrompt>;
+    /// Render one borrowed chat request into text or token IDs.
+    fn render(&self, request: RenderRequest<'_>) -> Result<RenderedPrompt>;
 }

 /// Shared trait-object form of [`ChatRenderer`].
@@ -49,7 +74,7 @@ pub type DynChatRenderer = Arc<dyn ChatRenderer>;
 /// using the provided defaults as the base.
 pub(crate) fn effective_template_kwargs(
    default_template_kwargs: &HashMap<String, Value>,
-    request: &ChatRequest,
+    request: &RenderRequest<'_>,
 ) -> HashMap<String, Value> {
    let mut kwargs = default_template_kwargs.clone();
    kwargs.extend(request.chat_options.template_kwargs.clone());
@@ -71,6 +96,6 @@ pub(crate) fn effective_template_kwargs(
 }

 /// Extract the effective chat-template kwargs visible to the renderer from the request.
-pub(crate) fn request_template_kwargs(request: &ChatRequest) -> HashMap<String, Value> {
+pub(crate) fn request_template_kwargs(request: &RenderRequest<'_>) -> HashMap<String, Value> {
    effective_template_kwargs(&HashMap::new(), request)
 }
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+use serde_json::Value;
+use vllm_chat_types::{
+    ChatMessage, ChatOptions, ChatRole, ChatToolChoice, GenerationPromptMode, Tool,
+};
+
+use crate::{Error, Result};
+
+/// Borrowed input consumed by one chat renderer.
+///
+/// This view contains only chat-domain values that affect prompt construction.
+/// Serving, sampling, decoding, scheduling, and engine metadata stay with the
+/// caller.
+#[derive(Debug, Clone, Copy)]
+pub struct RenderRequest<'a> {
+    /// Ordered chat history to render.
+    pub messages: &'a [ChatMessage],
+    /// Chat-template and generation-prompt controls.
+    pub chat_options: &'a ChatOptions,
+    /// Request-level tools available to the model.
+    pub tools: &'a [Tool],
+    /// Tool-choice behavior used to decide whether tools are exposed.
+    pub tool_choice: &'a ChatToolChoice,
+    /// Optional retrieval documents exposed to HF chat templates.
+    pub documents: Option<&'a [Value]>,
+}
+
+impl RenderRequest<'_> {
+    /// Validate renderer-owned request invariants.
+    pub fn validate(&self) -> Result<()> {
+        if self.messages.is_empty() {
+            return Err(Error::EmptyMessages);
+        }
+        match (
+            self.chat_options.generation_prompt_mode,
+            self.messages.last().map(ChatMessage::role),
+        ) {
+            (GenerationPromptMode::ContinueFinalAssistant, Some(ChatRole::Assistant)) => {}
+            (GenerationPromptMode::ContinueFinalAssistant, _) => {
+                return Err(Error::ContinueFinalAssistantWithoutFinalAssistant);
+            }
+            (GenerationPromptMode::NoGenerationPrompt, _)
+            | (GenerationPromptMode::StartNewAssistant, _) => {}
+        }
+        Ok(())
+    }
+
+    /// Return whether any message contains multimodal content.
+    pub fn has_multimodal(&self) -> bool {
+        self.messages.iter().any(ChatMessage::has_multimodal)
+    }
+
+    /// Return whether request-level tools should be exposed to the renderer.
+    pub fn tool_parsing_enabled(&self) -> bool {
+        !matches!(self.tool_choice, ChatToolChoice::None) && !self.tools.is_empty()
+    }
+
+    /// Return the request-level thinking toggle when explicitly requested.
+    ///
+    /// The `thinking` and `enable_thinking` kwargs must be booleans when
+    /// present and must carry the same value when both are set.
+    pub fn enable_thinking(&self) -> Result<Option<bool>> {
+        let thinking = self.parse_template_bool("thinking")?;
+        let enable_thinking = self.parse_template_bool("enable_thinking")?;
+
+        match (thinking, enable_thinking) {
+            (None, None) => Ok(None),
+            (Some(thinking), Some(enable_thinking)) if thinking != enable_thinking => {
+                Err(Error::ChatTemplate(
+                    "template kwargs `thinking` and `enable_thinking` must match when both are set"
+                        .to_string(),
+                ))
+            }
+            (Some(thinking), _) => Ok(Some(thinking)),
+            (None, Some(enable_thinking)) => Ok(Some(enable_thinking)),
+        }
+    }
+
+    /// Parse one optional boolean chat-template kwarg.
+    pub fn parse_template_bool(&self, key: &str) -> Result<Option<bool>> {
+        match self.chat_options.template_kwargs.get(key) {
+            None => Ok(None),
+            Some(Value::Bool(value)) => Ok(Some(*value)),
+            Some(other) => Err(Error::ChatTemplate(format!(
+                "template kwarg `{key}` must be a boolean, got {other}"
+            ))),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::json;
+
+    use super::RenderRequest;
+    use crate::{ChatMessage, ChatToolChoice, Error, GenerationPromptMode, TestRenderRequest};
+
+    #[test]
+    fn rejects_empty_message_history() {
+        let mut request = TestRenderRequest::for_test();
+        request.messages.clear();
+
+        assert!(matches!(
+            request.as_request().validate(),
+            Err(Error::EmptyMessages)
+        ));
+    }
+
+    #[test]
+    fn continuation_requires_a_final_assistant_message() {
+        let mut request = TestRenderRequest::for_test();
+        request.chat_options.generation_prompt_mode = GenerationPromptMode::ContinueFinalAssistant;
+
+        assert!(matches!(
+            request.as_request().validate(),
+            Err(Error::ContinueFinalAssistantWithoutFinalAssistant)
+        ));
+
+        request.messages.push(ChatMessage::assistant_text("partial"));
+        request.as_request().validate().unwrap();
+    }
+
+    #[test]
+    fn thinking_aliases_must_match() {
+        let mut request = TestRenderRequest::for_test();
+        request
+            .chat_options
+            .template_kwargs
+            .insert("thinking".to_string(), json!(false));
+        request
+            .chat_options
+            .template_kwargs
+            .insert("enable_thinking".to_string(), json!(true));
+
+        assert!(matches!(
+            request.as_request().enable_thinking(),
+            Err(Error::ChatTemplate(message))
+                if message.contains("`thinking` and `enable_thinking` must match")
+        ));
+    }
+
+    #[test]
+    fn tool_exposure_requires_tools_and_an_enabled_choice() {
+        let mut request = TestRenderRequest::for_test();
+        assert!(!request.as_request().tool_parsing_enabled());
+
+        request.tools.push(crate::Tool {
+            name: "lookup".to_string(),
+            description: None,
+            parameters: json!({"type": "object"}),
+            strict: None,
+        });
+        assert!(!request.as_request().tool_parsing_enabled());
+
+        request.tool_choice = ChatToolChoice::Auto;
+        assert!(request.as_request().tool_parsing_enabled());
+    }
+
+    #[test]
+    fn borrowed_request_carries_optional_documents() {
+        let mut request = TestRenderRequest::for_test();
+        request.documents = Some(vec![json!({"title": "doc"})]);
+
+        let borrowed: RenderRequest<'_> = request.as_request();
+        assert_eq!(
+            borrowed.documents.unwrap(),
+            request.documents.as_deref().unwrap()
+        );
+    }
+}
@@ -29,13 +29,21 @@ pub enum RendererSelection {
 }

 impl RendererSelection {
+    /// CLI/config literal for automatic renderer selection.
    pub const AUTO_LITERAL: &str = "auto";
+    /// CLI/config literal and model type for DeepSeek V3.2.
    pub const DEEPSEEK_V32_LITERAL: &str = "deepseek_v32";
+    /// CLI/config literal and model type for DeepSeek V4.
    pub const DEEPSEEK_V4_LITERAL: &str = "deepseek_v4";
+    /// Hugging Face model type routed to the Harmony renderer.
    pub const GPT_OSS_MODEL_TYPE: &str = "gpt_oss";
+    /// CLI/config literal for the Harmony renderer.
    pub const HARMONY_LITERAL: &str = "harmony";
+    /// CLI/config literal for the Hugging Face renderer.
    pub const HF_LITERAL: &str = "hf";
+    /// CLI/config literal for the Inkling renderer.
    pub const INKLING_LITERAL: &str = "inkling";
+    /// Hugging Face model type routed to the Inkling renderer.
    pub const INKLING_MODEL_TYPE: &str = "inkling_mm_model";

    /// Resolve the renderer selection using the given model type string, if
@@ -7,13 +7,45 @@ use std::path::Path;
 use serde::Deserialize;
 use serde_json::Value;

-use crate::event::{AssistantContentBlock, AssistantToolCall};
-use crate::request::{
-    ChatContent, ChatContentPart, ChatMessage, ChatRequest, ChatTool, ChatToolChoice,
-    GenerationPromptMode, ReasoningEffort,
+use crate::{
+    AssistantContentBlock, AssistantToolCall, ChatContent, ChatContentPart, ChatMessage,
+    ChatOptions, ChatRole, ChatToolChoice, GenerationPromptMode, ReasoningEffort, RenderRequest,
+    Tool,
 };

-/// Options for constructing a [`ChatRequest`] from a fixture file.
+/// Owned test input used to construct borrowed [`RenderRequest`] values.
+#[derive(Debug, Clone)]
+pub(crate) struct TestRenderRequest {
+    pub messages: Vec<ChatMessage>,
+    pub chat_options: ChatOptions,
+    pub tools: Vec<Tool>,
+    pub tool_choice: ChatToolChoice,
+    pub documents: Option<Vec<Value>>,
+}
+
+impl TestRenderRequest {
+    pub(crate) fn for_test() -> Self {
+        Self {
+            messages: vec![ChatMessage::text(ChatRole::User, "test")],
+            chat_options: ChatOptions::default(),
+            tools: Vec::new(),
+            tool_choice: ChatToolChoice::None,
+            documents: None,
+        }
+    }
+
+    pub(crate) fn as_request(&self) -> RenderRequest<'_> {
+        RenderRequest {
+            messages: &self.messages,
+            chat_options: &self.chat_options,
+            tools: &self.tools,
+            tool_choice: &self.tool_choice,
+            documents: self.documents.as_deref(),
+        }
+    }
+}
+
+/// Options for constructing a [`TestRenderRequest`] from a fixture file.
 #[derive(Debug, Clone, Copy)]
 pub(crate) struct FixtureRequestOptions {
    /// Whether to set the template kwarg `[enable_]thinking=true`.
@@ -23,9 +55,11 @@ pub(crate) struct FixtureRequestOptions {
    pub no_generation_prompt_when_last_assistant: bool,
 }

-/// Read a fixture file from the given path and convert it into a [`ChatRequest`]
-/// using the provided options.
-pub(crate) fn fixture_chat_request(path: &Path, options: FixtureRequestOptions) -> ChatRequest {
+/// Read a fixture file and convert it into an owned renderer test request.
+pub(crate) fn fixture_chat_request(
+    path: &Path,
+    options: FixtureRequestOptions,
+) -> TestRenderRequest {
    let fixture = fs::read_to_string(path).unwrap();
    let fixture: FixtureFile = serde_json::from_str(&fixture).unwrap();
    fixture.into_request().into_chat_request(options)
@@ -135,9 +169,8 @@ struct FixtureToolCallFunction {
 }

 impl FixtureRequest {
-    fn into_chat_request(self, options: FixtureRequestOptions) -> ChatRequest {
-        let mut request = ChatRequest {
-            request_id: "renderer-fixture".to_string(),
+    fn into_chat_request(self, options: FixtureRequestOptions) -> TestRenderRequest {
+        let mut request = TestRenderRequest {
            messages: self
                .messages
                .into_iter()
@@ -150,7 +183,7 @@ impl FixtureRequest {
            } else {
                ChatToolChoice::Auto
            },
-            ..ChatRequest::for_test()
+            ..TestRenderRequest::for_test()
        };

        if options.no_generation_prompt_when_last_assistant
@@ -243,10 +276,10 @@ fn to_chat_content(content: FixtureContent) -> ChatContent {
    }
 }

-fn to_chat_tools(tools: &[FixtureTool]) -> Vec<ChatTool> {
+fn to_chat_tools(tools: &[FixtureTool]) -> Vec<Tool> {
    tools
        .iter()
-        .map(|tool| ChatTool {
+        .map(|tool| Tool {
            name: tool.function.name.clone(),
            description: tool.function.description.clone(),
            parameters: tool.function.parameters.clone(),
@@ -0,0 +1,6 @@
+# vLLM Example Templates
+
+These fixtures are copied from `vllm/examples/`.
+
+They are used by `src/chat-renderer/src/hf/format.rs` tests to keep
+chat-template content-format detection aligned with Python vLLM behavior.
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Bugen Zhao	b6db7065e6	minor refactor Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-24 09:43:05 +00:00
Bugen Zhao	a2bedb84b6	do not take &&, better error categorization Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-24 09:26:14 +00:00
Bugen Zhao	21d8aa7ecc	simplify Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-24 09:22:01 +00:00
Bugen Zhao	4288f3d3a5	extract chat-renderer crate Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-24 09:10:46 +00:00
Bugen Zhao	97cbc212b4	[Rust Frontend] Extract shared chat types into `vllm-chat-types` Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-23 10:16:32 +00:00