feat: enable --dyn-reasoning-parser flag to set reasoning parser for vllm deployments (#2700)

Signed-off-by: Krishnan Prashanth <kprashanth@nvidia.com>

Author: nachiketb-nvidia

components/backends/vllm/src/dynamo/vllm/args.py

@@ -124,7 +124,7 @@ def parse_args() -> Config:
         "--dyn-reasoning-parser",
         type=str,
         default=None,
-        help="Reasoning parser name for the model.",
+        help="Reasoning parser name for the model. Available options: 'basic', 'deepseek_r1', 'gpt_oss'.",
     )
 
     parser = AsyncEngineArgs.add_cli_args(parser)

lib/llm/src/engines.rs

@@ -14,6 +14,7 @@ use dynamo_runtime::pipeline::{Error, ManyOut, SingleIn};
 use dynamo_runtime::protocols::annotated::Annotated;
 
 use crate::backend::ExecutionContext;
+use crate::local_model::runtime_config;
 use crate::preprocessor::PreprocessedRequest;
 use crate::protocols::common::llm_backend::LLMEngineOutput;
 use crate::protocols::openai::{
@@ -183,7 +184,7 @@ impl
         incoming_request: SingleIn<NvCreateChatCompletionRequest>,
     ) -> Result<ManyOut<Annotated<NvCreateChatCompletionStreamResponse>>, Error> {
         let (request, context) = incoming_request.transfer(());
-        let mut deltas = request.response_generator();
+        let mut deltas = request.response_generator(runtime_config::ModelRuntimeConfig::default());
         let ctx = context.context();
         let req = request.inner.messages.into_iter().next_back().unwrap();
 

lib/llm/src/local_model.rs

@@ -209,6 +209,7 @@ impl LocalModelBuilder {
             );
             card.migration_limit = self.migration_limit;
             card.user_data = self.user_data.take();
+            card.runtime_config = self.runtime_config.clone();
 
             return Ok(LocalModel {
                 card,
@@ -287,6 +288,7 @@ impl LocalModelBuilder {
 
         card.migration_limit = self.migration_limit;
         card.user_data = self.user_data.take();
+        card.runtime_config = self.runtime_config.clone();
 
         Ok(LocalModel {
             card,

lib/llm/src/model_card.rs

@@ -19,6 +19,7 @@ use std::path::{Path, PathBuf};
 use std::sync::Arc;
 use std::time::Duration;
 
+use crate::local_model::runtime_config::ModelRuntimeConfig;
 use anyhow::{Context, Result};
 use derive_builder::Builder;
 use dynamo_runtime::{slug::Slug, storage::key_value_store::Versioned, transports::nats};
@@ -152,6 +153,9 @@ pub struct ModelDeploymentCard {
     /// User-defined metadata for custom worker behavior
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub user_data: Option<serde_json::Value>,
+
+    #[serde(default)]
+    pub runtime_config: ModelRuntimeConfig,
 }
 
 impl ModelDeploymentCard {
@@ -488,6 +492,7 @@ impl ModelDeploymentCard {
             kv_cache_block_size: 0,
             migration_limit: 0,
             user_data: None,
+            runtime_config: ModelRuntimeConfig::default(),
         })
     }
 
@@ -556,6 +561,7 @@ impl ModelDeploymentCard {
             kv_cache_block_size: 0, // set later
             migration_limit: 0,
             user_data: None,
+            runtime_config: ModelRuntimeConfig::default(),
         })
     }
 }

lib/llm/src/preprocessor.rs

@@ -22,6 +22,7 @@ use rayon::iter::{IntoParallelRefIterator, ParallelIterator};
 use std::{collections::HashMap, sync::Arc};
 use tracing;
 
+use crate::local_model::runtime_config::ModelRuntimeConfig;
 use crate::model_card::{ModelDeploymentCard, ModelInfo, TokenizerKind};
 use crate::preprocessor::prompt::OAIChatLikeRequest;
 use crate::tokenizers::Encoding;
@@ -94,6 +95,7 @@ pub struct OpenAIPreprocessor {
     formatter: Arc<dyn OAIPromptFormatter>,
     tokenizer: Arc<dyn Tokenizer>,
     model_info: Arc<dyn ModelInfo>,
+    runtime_config: ModelRuntimeConfig,
 }
 
 impl OpenAIPreprocessor {
@@ -121,11 +123,14 @@ impl OpenAIPreprocessor {
         };
         let model_info = model_info.get_model_info().await?;
 
+        let runtime_config = mdc.runtime_config.clone();
+
         Ok(Arc::new(Self {
             formatter,
             tokenizer,
             model_info,
             mdcsum,
+            runtime_config,
         }))
     }
 
@@ -494,7 +499,7 @@ impl
         let (request, context) = request.into_parts();
 
         // create a response generator
-        let response_generator = request.response_generator();
+        let response_generator = request.response_generator(self.runtime_config.clone());
         let mut response_generator = Box::new(response_generator);
 
         // convert the chat completion request to a common completion request

lib/llm/src/protocols/openai/chat_completions/delta.rs

@@ -5,6 +5,7 @@ use dynamo_parsers::{ParserResult, ReasoningParser, ReasoningParserType, Reasoni
 
 use super::{NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse};
 use crate::{
+    local_model::runtime_config,
     protocols::common::{self},
     types::TokenIdType,
 };
@@ -15,11 +16,15 @@ impl NvCreateChatCompletionRequest {
     ///
     /// # Returns
     /// * [`DeltaGenerator`] configured with model name and response options.
-    pub fn response_generator(&self) -> DeltaGenerator {
+    pub fn response_generator(
+        &self,
+        runtime_config: runtime_config::ModelRuntimeConfig,
+    ) -> DeltaGenerator {
         let options = DeltaGeneratorOptions {
             enable_usage: true,
             enable_logprobs: self.inner.logprobs.unwrap_or(false)
                 || self.inner.top_logprobs.unwrap_or(0) > 0,
+            runtime_config,
         };
 
         DeltaGenerator::new(self.inner.model.clone(), options)
@@ -33,6 +38,8 @@ pub struct DeltaGeneratorOptions {
     pub enable_usage: bool,
     /// Determines whether log probabilities should be included in the response.
     pub enable_logprobs: bool,
+
+    pub runtime_config: runtime_config::ModelRuntimeConfig,
 }
 
 /// Generates incremental chat completion responses in a streaming fashion.
@@ -92,10 +99,14 @@ impl DeltaGenerator {
         // This is hardcoded for now, but can be made configurable later.
         // TODO: Make parser type configurable once front-end integration is determined
         // Change to GptOss to test GptOSS parser
-        let reasoning_parser_type = ReasoningParserType::Basic;
-
         // Reasoning parser wrapper
-        let reasoning_parser = reasoning_parser_type.get_reasoning_parser();
+        let reasoning_parser = ReasoningParserType::get_reasoning_parser_from_name(
+            options
+                .runtime_config
+                .reasoning_parser
+                .as_deref()
+                .unwrap_or("basic"),
+        );
 
         Self {
             id: format!("chatcmpl-{}", uuid::Uuid::new_v4()),

lib/llm/tests/http-service.rs

@@ -16,17 +16,6 @@
 use anyhow::Error;
 use async_stream::stream;
 use dynamo_async_openai::config::OpenAIConfig;
-use dynamo_llm::http::{
-    client::{
-        GenericBYOTClient, HttpClientConfig, HttpRequestContext, NvCustomClient, PureOpenAIClient,
-    },
-    service::{
-        Metrics,
-        error::HttpError,
-        metrics::{Endpoint, FRONTEND_METRIC_PREFIX, RequestType, Status},
-        service_v2::HttpService,
-    },
-};
 use dynamo_llm::protocols::{
     Annotated,
     codec::SseLineCodec,
@@ -36,6 +25,21 @@ use dynamo_llm::protocols::{
         completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
     },
 };
+use dynamo_llm::{
+    http::{
+        client::{
+            GenericBYOTClient, HttpClientConfig, HttpRequestContext, NvCustomClient,
+            PureOpenAIClient,
+        },
+        service::{
+            Metrics,
+            error::HttpError,
+            metrics::{Endpoint, FRONTEND_METRIC_PREFIX, RequestType, Status},
+            service_v2::HttpService,
+        },
+    },
+    local_model::runtime_config,
+};
 use dynamo_runtime::{
     CancellationToken,
     engine::AsyncEngineContext,
@@ -95,7 +99,8 @@ impl
         let max_tokens = request.inner.max_tokens.unwrap_or(0) as u64;
 
         // let generator = NvCreateChatCompletionStreamResponse::generator(request.model.clone());
-        let mut generator = request.response_generator();
+        let mut generator =
+            request.response_generator(runtime_config::ModelRuntimeConfig::default());
 
         let stream = stream! {
             tokio::time::sleep(std::time::Duration::from_millis(max_tokens)).await;

lib/parsers/src/reasoning/mod.rs

@@ -115,4 +115,20 @@ impl ReasoningParserType {
             },
         }
     }
+
+    pub fn get_reasoning_parser_from_name(name: &str) -> ReasoningParserWrapper {
+        tracing::debug!("Selected reasoning parser: {}", name);
+        match name.to_lowercase().as_str() {
+            "deepseek_r1" => Self::DeepseekR1.get_reasoning_parser(),
+            "basic" => Self::Basic.get_reasoning_parser(),
+            "gpt_oss" => Self::GptOss.get_reasoning_parser(),
+            _ => {
+                tracing::warn!(
+                    "Unknown reasoning parser type '{}', falling back to Basic Reasoning Parser",
+                    name
+                );
+                Self::Basic.get_reasoning_parser()
+            }
+        }
+    }
 }