[Rust Frontend] Correctly parse whitespace framing in model output (#55417)
Signed-off-by: Bugen Zhao <i@bugenzhao.com> Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
@@ -416,6 +416,6 @@ mod tests {
|
||||
)
|
||||
.unwrap_err();
|
||||
|
||||
expect_test::expect!["reasoning parser `definitely_missing_reasoning_parser` is not registered (choose from: cohere_cmd, deepseek_r1, deepseek_v3, deepseek_v4, gemma4, glm45, hy_v3, hy_v4, inkling, kimi, kimi_k2, kimi_k3, minimax_m2, minimax_m3, nemotron_v3, qwen3, seed_oss, step3, step3p5)"].assert_eq(&error.to_report_string());
|
||||
expect_test::expect!["reasoning parser `definitely_missing_reasoning_parser` is not registered (choose from: cohere_cmd, deepseek_r1, deepseek_v3, deepseek_v4, gemma4, glm45, glm47, hy_v3, hy_v4, inkling, kimi, kimi_k2, kimi_k3, minimax_m2, minimax_m3, nemotron_v3, qwen3, seed_oss, step3, step3p5)"].assert_eq(&error.to_report_string());
|
||||
}
|
||||
}
|
||||
|
||||
@@ -7,10 +7,10 @@ use std::sync::{Arc, LazyLock};
|
||||
|
||||
pub use vllm_parser::reasoning::{
|
||||
CohereCmdReasoningParser, DeepSeekR1ReasoningParser, DeepSeekV3ReasoningParser,
|
||||
DeepSeekV4ReasoningParser, Glm45ReasoningParser, KimiK2ReasoningParser, KimiReasoningParser,
|
||||
MiniMaxM2ReasoningParser, MiniMaxM3ReasoningParser, NemotronV3ReasoningParser,
|
||||
Qwen3ReasoningParser, ReasoningDelta, ReasoningError, ReasoningParser, SeedOssReasoningParser,
|
||||
Step3ReasoningParser, Step3p5ReasoningParser,
|
||||
DeepSeekV4ReasoningParser, Glm45ReasoningParser, Glm47ReasoningParser, KimiK2ReasoningParser,
|
||||
KimiReasoningParser, MiniMaxM2ReasoningParser, MiniMaxM3ReasoningParser,
|
||||
NemotronV3ReasoningParser, Qwen3ReasoningParser, ReasoningDelta, ReasoningError,
|
||||
ReasoningParser, SeedOssReasoningParser, Step3ReasoningParser, Step3p5ReasoningParser,
|
||||
};
|
||||
use vllm_tokenizer::DynTokenizer;
|
||||
|
||||
@@ -25,6 +25,7 @@ pub mod names {
|
||||
pub const GEMMA4: &str = "gemma4";
|
||||
pub const INKLING: &str = "inkling";
|
||||
pub const GLM45: &str = "glm45";
|
||||
pub const GLM47: &str = "glm47";
|
||||
pub const HY_V3: &str = "hy_v3";
|
||||
pub const HY_V4: &str = "hy_v4";
|
||||
pub const KIMI: &str = "kimi";
|
||||
@@ -69,6 +70,7 @@ impl ReasoningParserFactory {
|
||||
.register_unified_dummy(names::GEMMA4)
|
||||
.register_unified_dummy(names::INKLING)
|
||||
.register_parser::<Glm45ReasoningParser>(names::GLM45)
|
||||
.register_parser::<Glm47ReasoningParser>(names::GLM47)
|
||||
.register_unified_dummy(names::HY_V3)
|
||||
.register_unified_dummy(names::HY_V4)
|
||||
.register_parser::<KimiReasoningParser>(names::KIMI)
|
||||
@@ -91,8 +93,8 @@ impl ReasoningParserFactory {
|
||||
.register_pattern("gemma4", names::GEMMA4)
|
||||
.register_pattern("qwq", names::DEEPSEEK_R1)
|
||||
.register_pattern("qwen3", names::QWEN3)
|
||||
.register_pattern("glm-5", names::GLM45)
|
||||
.register_pattern("glm-4.7", names::GLM45)
|
||||
.register_pattern("glm-5", names::GLM47)
|
||||
.register_pattern("glm-4.7", names::GLM47)
|
||||
.register_pattern("glm-4.6", names::GLM45)
|
||||
.register_pattern("glm-4.5", names::GLM45)
|
||||
.register_pattern("hy3", names::HY_V3)
|
||||
|
||||
@@ -25,7 +25,7 @@ fn factory_contains_and_lists_registered_parsers() {
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn factory_resolves_deepseek_v4_to_qwen3_alias() {
|
||||
fn factory_resolves_deepseek_v4() {
|
||||
let factory = ReasoningParserFactory::new();
|
||||
assert_eq!(
|
||||
factory.resolve_name_for_model("deepseek-ai/DeepSeek-V4"),
|
||||
@@ -112,3 +112,14 @@ fn factory_rejects_unknown_parser_names() {
|
||||
};
|
||||
assert!(error.to_string().contains("choose from"));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn factory_distinguishes_glm_reasoning_framing() {
|
||||
let factory = ReasoningParserFactory::new();
|
||||
for model in ["zai-org/GLM-4.5", "zai-org/GLM-4.6"] {
|
||||
assert_eq!(factory.resolve_name_for_model(model), Some(names::GLM45));
|
||||
}
|
||||
for model in ["zai-org/GLM-4.7-Flash", "zai-org/GLM-5.2-FP8"] {
|
||||
assert_eq!(factory.resolve_name_for_model(model), Some(names::GLM47));
|
||||
}
|
||||
}
|
||||
|
||||
@@ -445,10 +445,12 @@ async fn run_roundtrip_reasoning_and_content_inner(
|
||||
let result = run_roundtrip(case, backends, &request, assistant).await?;
|
||||
|
||||
assert_eq!(
|
||||
result.parsed_message.reasoning().as_deref().map(str::trim),
|
||||
effective_thinking.then_some(expected_reasoning)
|
||||
result.parsed_message.reasoning().as_deref(),
|
||||
effective_thinking.then_some(expected_reasoning),
|
||||
"parsed message: {:#?}",
|
||||
result.parsed_message
|
||||
);
|
||||
assert_eq!(result.parsed_message.text().trim(), expected_text);
|
||||
assert_eq!(result.parsed_message.text(), expected_text);
|
||||
assert_eq!(result.parsed_message.tool_calls().count(), 0);
|
||||
|
||||
assert_eq!(
|
||||
@@ -508,10 +510,12 @@ async fn run_roundtrip_tool_call_mix(
|
||||
.await?;
|
||||
|
||||
assert_eq!(
|
||||
result.parsed_message.reasoning().as_deref().map(str::trim),
|
||||
Some(expected_reasoning)
|
||||
result.parsed_message.reasoning().as_deref(),
|
||||
Some(expected_reasoning),
|
||||
"parsed message: {:#?}",
|
||||
result.parsed_message
|
||||
);
|
||||
assert_eq!(result.parsed_message.text().trim(), expected_text);
|
||||
assert_eq!(result.parsed_message.text(), expected_text);
|
||||
|
||||
let tool_calls = result.parsed_message.tool_calls().collect::<Vec<_>>();
|
||||
assert_eq!(
|
||||
@@ -714,11 +718,7 @@ async fn parse_completion(
|
||||
|
||||
while let Some(event) = events.next().await {
|
||||
if let ChatEvent::Done { message, .. } = event? {
|
||||
// TODO: currently our parsers are not very strict about preserving or trimming
|
||||
// whitespace, so we trim here to avoid roundtrip failures due to
|
||||
// insignificant whitespace differences. However, this may hurt token-level
|
||||
// fidelity so we should consider improving them.
|
||||
return Ok(message.trim());
|
||||
return Ok(message);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user