refactor(2/3): rename dynamo-async-openai to dynamo-protocols (#7565)

lib/llm/src/preprocessor/speculative_prefill.rs

@@ -12,7 +12,7 @@ use std::pin::Pin;
 use std::sync::Arc;
 
 use anyhow::Result;
-use dynamo_async_openai::types::{
+use dynamo_protocols::types::{
     ChatCompletionMessageContent, ChatCompletionRequestAssistantMessage,
     ChatCompletionRequestAssistantMessageContent, ChatCompletionRequestMessage,
 };

lib/llm/src/protocols/anthropic/stream_converter.rs

@@ -10,7 +10,7 @@
 use std::collections::HashSet;
 
 use axum::response::sse::Event;
-use dynamo_async_openai::types::ChatCompletionMessageContent;
+use dynamo_protocols::types::ChatCompletionMessageContent;
 use uuid::Uuid;
 
 use super::types::{
@@ -136,17 +136,15 @@ impl AnthropicStreamConverter {
             // Track finish reason
             if let Some(ref fr) = choice.finish_reason {
                 self.stop_reason = Some(match fr {
-                    dynamo_async_openai::types::FinishReason::Stop => AnthropicStopReason::EndTurn,
-                    dynamo_async_openai::types::FinishReason::Length => {
-                        AnthropicStopReason::MaxTokens
-                    }
-                    dynamo_async_openai::types::FinishReason::ToolCalls => {
+                    dynamo_protocols::types::FinishReason::Stop => AnthropicStopReason::EndTurn,
+                    dynamo_protocols::types::FinishReason::Length => AnthropicStopReason::MaxTokens,
+                    dynamo_protocols::types::FinishReason::ToolCalls => {
                         AnthropicStopReason::ToolUse
                     }
-                    dynamo_async_openai::types::FinishReason::ContentFilter => {
+                    dynamo_protocols::types::FinishReason::ContentFilter => {
                         AnthropicStopReason::EndTurn
                     }
-                    dynamo_async_openai::types::FinishReason::FunctionCall => {
+                    dynamo_protocols::types::FinishReason::FunctionCall => {
                         AnthropicStopReason::ToolUse
                     }
                 });
@@ -478,17 +476,15 @@ impl AnthropicStreamConverter {
 
             if let Some(ref fr) = choice.finish_reason {
                 self.stop_reason = Some(match fr {
-                    dynamo_async_openai::types::FinishReason::Stop => AnthropicStopReason::EndTurn,
-                    dynamo_async_openai::types::FinishReason::Length => {
-                        AnthropicStopReason::MaxTokens
-                    }
-                    dynamo_async_openai::types::FinishReason::ToolCalls => {
+                    dynamo_protocols::types::FinishReason::Stop => AnthropicStopReason::EndTurn,
+                    dynamo_protocols::types::FinishReason::Length => AnthropicStopReason::MaxTokens,
+                    dynamo_protocols::types::FinishReason::ToolCalls => {
                         AnthropicStopReason::ToolUse
                     }
-                    dynamo_async_openai::types::FinishReason::ContentFilter => {
+                    dynamo_protocols::types::FinishReason::ContentFilter => {
                         AnthropicStopReason::EndTurn
                     }
-                    dynamo_async_openai::types::FinishReason::FunctionCall => {
+                    dynamo_protocols::types::FinishReason::FunctionCall => {
                         AnthropicStopReason::ToolUse
                     }
                 });
@@ -734,7 +730,7 @@ impl AnthropicStreamConverter {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use dynamo_async_openai::types::{
+    use dynamo_protocols::types::{
         ChatChoiceStream, ChatCompletionMessageContent, ChatCompletionMessageToolCallChunk,
         ChatCompletionStreamResponseDelta, ChatCompletionToolType, FunctionCallStream,
     };
@@ -742,7 +738,7 @@ mod tests {
     fn text_chunk(text: &str) -> NvCreateChatCompletionStreamResponse {
         #[allow(deprecated)]
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: "chat-1".into(),
                 choices: vec![ChatChoiceStream {
                     index: 0,
@@ -777,7 +773,7 @@ mod tests {
     ) -> NvCreateChatCompletionStreamResponse {
         #[allow(deprecated)]
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: "chat-1".into(),
                 choices: vec![ChatChoiceStream {
                     index: 0,
@@ -932,7 +928,7 @@ mod tests {
     fn reasoning_chunk(text: &str) -> NvCreateChatCompletionStreamResponse {
         #[allow(deprecated)]
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: "chat-1".into(),
                 choices: vec![ChatChoiceStream {
                     index: 0,

lib/llm/src/protocols/anthropic/types.rs

@@ -3,15 +3,15 @@
 
 //! Anthropic Messages API conversion logic.
 //!
-//! Pure protocol types live in `dynamo_async_openai::types::anthropic`.
+//! Pure protocol types live in `dynamo_protocols::types::anthropic`.
 //! This module provides bidirectional conversion to/from the internal
 //! chat completions format used by the Dynamo engine.
 
 // Re-export all pure Anthropic protocol types so existing `use crate::protocols::anthropic::*`
 // continues to work throughout dynamo-llm.
-pub use dynamo_async_openai::types::anthropic::*;
+pub use dynamo_protocols::types::anthropic::*;
 
-use dynamo_async_openai::types::{
+use dynamo_protocols::types::{
     ChatCompletionMessageToolCall, ChatCompletionNamedToolChoice,
     ChatCompletionRequestAssistantMessage, ChatCompletionRequestAssistantMessageContent,
     ChatCompletionRequestMessage, ChatCompletionRequestMessageContentPartImage,
@@ -96,10 +96,10 @@ impl TryFrom<AnthropicCreateMessageRequest> for NvCreateChatCompletionRequest {
         // Convert stop_sequences -> stop
         let stop = req
             .stop_sequences
-            .map(dynamo_async_openai::types::Stop::StringArray);
+            .map(dynamo_protocols::types::Stop::StringArray);
 
         Ok(NvCreateChatCompletionRequest {
-            inner: dynamo_async_openai::types::CreateChatCompletionRequest {
+            inner: dynamo_protocols::types::CreateChatCompletionRequest {
                 messages,
                 model: req.model,
                 temperature: req.temperature,
@@ -109,7 +109,7 @@ impl TryFrom<AnthropicCreateMessageRequest> for NvCreateChatCompletionRequest {
                 tools,
                 tool_choice,
                 stream: Some(true), // Always stream internally
-                stream_options: Some(dynamo_async_openai::types::ChatCompletionStreamOptions {
+                stream_options: Some(dynamo_protocols::types::ChatCompletionStreamOptions {
                     include_usage: true,
                     continuous_usage_stats: false,
                 }),
@@ -350,7 +350,7 @@ fn convert_assistant_blocks(
                 tool_calls.push(ChatCompletionMessageToolCall {
                     id: id.clone(),
                     r#type: ChatCompletionToolType::Function,
-                    function: dynamo_async_openai::types::FunctionCall {
+                    function: dynamo_protocols::types::FunctionCall {
                         name: name.clone(),
                         arguments: serde_json::to_string(input).unwrap_or_default(),
                     },
@@ -487,11 +487,11 @@ pub fn chat_completion_to_anthropic_response(
     if let Some(choice) = choice {
         // Map finish_reason
         stop_reason = choice.finish_reason.map(|fr| match fr {
-            dynamo_async_openai::types::FinishReason::Stop => AnthropicStopReason::EndTurn,
-            dynamo_async_openai::types::FinishReason::Length => AnthropicStopReason::MaxTokens,
-            dynamo_async_openai::types::FinishReason::ToolCalls => AnthropicStopReason::ToolUse,
-            dynamo_async_openai::types::FinishReason::ContentFilter => AnthropicStopReason::EndTurn,
-            dynamo_async_openai::types::FinishReason::FunctionCall => AnthropicStopReason::ToolUse,
+            dynamo_protocols::types::FinishReason::Stop => AnthropicStopReason::EndTurn,
+            dynamo_protocols::types::FinishReason::Length => AnthropicStopReason::MaxTokens,
+            dynamo_protocols::types::FinishReason::ToolCalls => AnthropicStopReason::ToolUse,
+            dynamo_protocols::types::FinishReason::ContentFilter => AnthropicStopReason::EndTurn,
+            dynamo_protocols::types::FinishReason::FunctionCall => AnthropicStopReason::ToolUse,
         });
 
         // Extract tool calls
@@ -523,8 +523,8 @@ pub fn chat_completion_to_anthropic_response(
 
         // Extract text content
         let text = match choice.message.content {
-            Some(dynamo_async_openai::types::ChatCompletionMessageContent::Text(t)) => Some(t),
-            Some(dynamo_async_openai::types::ChatCompletionMessageContent::Parts(_)) => {
+            Some(dynamo_protocols::types::ChatCompletionMessageContent::Text(t)) => Some(t),
+            Some(dynamo_protocols::types::ChatCompletionMessageContent::Parts(_)) => {
                 tracing::warn!(
                     "Multimodal (Parts) content in chat completion response replaced with placeholder text in Anthropic conversion."
                 );
@@ -821,24 +821,22 @@ mod tests {
     #[test]
     fn test_chat_completion_to_anthropic_response() {
         let chat_resp = NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 id: "chatcmpl-xyz".into(),
-                choices: vec![dynamo_async_openai::types::ChatChoice {
+                choices: vec![dynamo_protocols::types::ChatChoice {
                     index: 0,
-                    message: dynamo_async_openai::types::ChatCompletionResponseMessage {
-                        content: Some(
-                            dynamo_async_openai::types::ChatCompletionMessageContent::Text(
+                    message: dynamo_protocols::types::ChatCompletionResponseMessage {
+                        content: Some(dynamo_protocols::types::ChatCompletionMessageContent::Text(
                             "Hello!".to_string(),
-                            ),
-                        ),
+                        )),
                         refusal: None,
                         tool_calls: None,
-                        role: dynamo_async_openai::types::Role::Assistant,
+                        role: dynamo_protocols::types::Role::Assistant,
                         function_call: None,
                         audio: None,
                         reasoning_content: None,
                     },
-                    finish_reason: Some(dynamo_async_openai::types::FinishReason::Stop),
+                    finish_reason: Some(dynamo_protocols::types::FinishReason::Stop),
                     stop_reason: None,
                     logprobs: None,
                 }],
@@ -847,7 +845,7 @@ mod tests {
                 service_tier: None,
                 system_fingerprint: None,
                 object: "chat.completion".to_string(),
-                usage: Some(dynamo_async_openai::types::CompletionUsage {
+                usage: Some(dynamo_protocols::types::CompletionUsage {
                     prompt_tokens: 10,
                     completion_tokens: 5,
                     total_tokens: 15,

lib/llm/src/protocols/common.rs

@@ -90,27 +90,27 @@ impl std::str::FromStr for FinishReason {
     }
 }
 
-impl From<FinishReason> for dynamo_async_openai::types::CompletionFinishReason {
+impl From<FinishReason> for dynamo_protocols::types::CompletionFinishReason {
     fn from(reason: FinishReason) -> Self {
         match reason {
             FinishReason::EoS | FinishReason::Stop | FinishReason::Cancelled => {
-                dynamo_async_openai::types::CompletionFinishReason::Stop
+                dynamo_protocols::types::CompletionFinishReason::Stop
             }
             FinishReason::ContentFilter => {
-                dynamo_async_openai::types::CompletionFinishReason::ContentFilter
+                dynamo_protocols::types::CompletionFinishReason::ContentFilter
             }
-            FinishReason::Length => dynamo_async_openai::types::CompletionFinishReason::Length,
-            FinishReason::Error(_) => dynamo_async_openai::types::CompletionFinishReason::Stop,
+            FinishReason::Length => dynamo_protocols::types::CompletionFinishReason::Length,
+            FinishReason::Error(_) => dynamo_protocols::types::CompletionFinishReason::Stop,
         }
     }
 }
 
-impl From<dynamo_async_openai::types::CompletionFinishReason> for FinishReason {
-    fn from(reason: dynamo_async_openai::types::CompletionFinishReason) -> Self {
+impl From<dynamo_protocols::types::CompletionFinishReason> for FinishReason {
+    fn from(reason: dynamo_protocols::types::CompletionFinishReason) -> Self {
         match reason {
-            dynamo_async_openai::types::CompletionFinishReason::Stop => FinishReason::Stop,
-            dynamo_async_openai::types::CompletionFinishReason::Length => FinishReason::Length,
-            dynamo_async_openai::types::CompletionFinishReason::ContentFilter => {
+            dynamo_protocols::types::CompletionFinishReason::Stop => FinishReason::Stop,
+            dynamo_protocols::types::CompletionFinishReason::Length => FinishReason::Length,
+            dynamo_protocols::types::CompletionFinishReason::ContentFilter => {
                 FinishReason::ContentFilter
             }
         }

lib/llm/src/protocols/common/llm_backend.rs

@@ -6,8 +6,8 @@ use serde::{Deserialize, Serialize};
 pub use super::FinishReason;
 pub use super::preprocessor::PreprocessedRequest;
 use crate::protocols::TokenIdType;
-use dynamo_async_openai::types::CompletionUsage;
-use dynamo_async_openai::types::StopReason;
+use dynamo_protocols::types::CompletionUsage;
+use dynamo_protocols::types::StopReason;
 use dynamo_runtime::error::DynamoError;
 use dynamo_runtime::protocols::maybe_error::MaybeError;
 

lib/llm/src/protocols/common/preprocessor.rs

@@ -86,7 +86,7 @@ pub struct PrefillResult {
     pub disaggregated_params: serde_json::Value,
     /// Prompt token details produced during prefill
     #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub prompt_tokens_details: Option<dynamo_async_openai::types::PromptTokensDetails>,
+    pub prompt_tokens_details: Option<dynamo_protocols::types::PromptTokensDetails>,
 }
 
 /// Optional multimodal routing-only data.

lib/llm/src/protocols/openai.rs

@@ -230,15 +230,15 @@ pub(crate) fn convert_backend_top_logprobs(
     selected_token: &str,
     selected_token_id: TokenIdType,
     selected_logprob: f32,
-) -> Vec<dynamo_async_openai::types::TopLogprobs> {
+) -> Vec<dynamo_protocols::types::TopLogprobs> {
     let mut found_selected = false;
-    let mut result: Vec<dynamo_async_openai::types::TopLogprobs> = top_lps
+    let mut result: Vec<dynamo_protocols::types::TopLogprobs> = top_lps
         .iter()
         .map(|top_lp| {
             let tok = top_lp.token.clone().unwrap_or_default();
             found_selected = found_selected || top_lp.token_id == selected_token_id;
             let bytes = top_lp.bytes.clone().or_else(|| token_to_utf8_bytes(&tok));
-            dynamo_async_openai::types::TopLogprobs {
+            dynamo_protocols::types::TopLogprobs {
                 token: tok,
                 logprob: top_lp.logprob as f32,
                 bytes,
@@ -247,7 +247,7 @@ pub(crate) fn convert_backend_top_logprobs(
         .collect();
 
     if !found_selected {
-        result.push(dynamo_async_openai::types::TopLogprobs {
+        result.push(dynamo_protocols::types::TopLogprobs {
             token: selected_token.to_string(),
             logprob: selected_logprob,
             bytes: token_to_utf8_bytes(selected_token),
@@ -277,7 +277,7 @@ pub trait DeltaGeneratorExt<ResponseType: Send + 'static + std::fmt::Debug>:
     fn is_continuous_usage_enabled(&self) -> bool;
 
     /// Get the current usage statistics with properly calculated total_tokens.
-    fn get_usage(&self) -> dynamo_async_openai::types::CompletionUsage;
+    fn get_usage(&self) -> dynamo_protocols::types::CompletionUsage;
 
     /// Returns the request tracker if available, for accessing worker timing metrics.
     fn tracker(&self) -> Option<std::sync::Arc<common::timing::RequestTracker>> {

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

@@ -35,7 +35,7 @@ pub use delta::DeltaGenerator;
 #[derive(ToSchema, Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvCreateChatCompletionRequest {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateChatCompletionRequest,
+    pub inner: dynamo_protocols::types::CreateChatCompletionRequest,
 
     #[serde(flatten, default)]
     pub common: CommonExt,
@@ -68,7 +68,7 @@ pub struct NvCreateChatCompletionRequest {
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
 pub struct NvCreateChatCompletionResponse {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateChatCompletionResponse,
+    pub inner: dynamo_protocols::types::CreateChatCompletionResponse,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub nvext: Option<serde_json::Value>,
 }
@@ -78,7 +78,7 @@ pub struct NvCreateChatCompletionResponse {
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
 pub struct NvCreateChatCompletionStreamResponse {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse,
+    pub inner: dynamo_protocols::types::CreateChatCompletionStreamResponse,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub nvext: Option<serde_json::Value>,
 }
@@ -202,7 +202,7 @@ impl CommonExtProvider for NvCreateChatCompletionRequest {
 
         // 2) OpenAI `response_format` (applies to assistant content, not tool calls)
         if let Some(response_format) = self.inner.response_format.as_ref() {
-            use dynamo_async_openai::types::ResponseFormat;
+            use dynamo_protocols::types::ResponseFormat;
             match response_format {
                 ResponseFormat::Text => {}
                 ResponseFormat::JsonObject => {
@@ -289,8 +289,8 @@ impl OpenAIStopConditionsProvider for NvCreateChatCompletionRequest {
     /// * `None` if no stop conditions are defined.
     fn get_stop(&self) -> Option<Vec<String>> {
         self.inner.stop.as_ref().map(|stop| match stop {
-            dynamo_async_openai::types::Stop::String(s) => vec![s.clone()],
-            dynamo_async_openai::types::Stop::StringArray(arr) => arr.clone(),
+            dynamo_protocols::types::Stop::String(s) => vec![s.clone()],
+            dynamo_protocols::types::Stop::StringArray(arr) => arr.clone(),
         })
     }
 

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

@@ -32,7 +32,7 @@ impl NvCreateChatCompletionRequest {
             // For non-streaming requests (stream=false), enable usage by default
             if self.inner.stream_options.is_none() {
                 self.inner.stream_options =
-                    Some(dynamo_async_openai::types::ChatCompletionStreamOptions {
+                    Some(dynamo_protocols::types::ChatCompletionStreamOptions {
                         include_usage: true,
                         continuous_usage_stats: false,
                     });
@@ -116,9 +116,9 @@ pub struct DeltaGenerator {
     /// Optional system fingerprint for version tracking.
     system_fingerprint: Option<String>,
     /// Optional service tier information for the response.
-    service_tier: Option<dynamo_async_openai::types::ServiceTierResponse>,
+    service_tier: Option<dynamo_protocols::types::ServiceTierResponse>,
     /// Tracks token usage for the completion request.
-    usage: dynamo_async_openai::types::CompletionUsage,
+    usage: dynamo_protocols::types::CompletionUsage,
     /// Counter tracking the number of messages issued.
     msg_counter: u64,
     /// Configuration options for response generation.
@@ -147,7 +147,7 @@ impl DeltaGenerator {
         // but this will not be an issue until 2106.
         let now: u32 = now.try_into().expect("timestamp exceeds u32::MAX");
 
-        let usage = dynamo_async_openai::types::CompletionUsage {
+        let usage = dynamo_protocols::types::CompletionUsage {
             prompt_tokens: 0,
             completion_tokens: 0,
             total_tokens: 0,
@@ -194,7 +194,7 @@ impl DeltaGenerator {
         token_ids: &[TokenIdType],
         logprobs: Option<common::llm_backend::LogProbs>,
         top_logprobs: Option<common::llm_backend::TopLogprobs>,
-    ) -> Option<dynamo_async_openai::types::ChatChoiceLogprobs> {
+    ) -> Option<dynamo_protocols::types::ChatChoiceLogprobs> {
         if !self.options.enable_logprobs || logprobs.is_none() {
             return None;
         }
@@ -216,7 +216,7 @@ impl DeltaGenerator {
                 .zip(top_logprobs)
                 .map(|(((t, tid), lp), top_lps)| {
                     let converted = convert_backend_top_logprobs(&top_lps, t, *tid, lp);
-                    dynamo_async_openai::types::ChatCompletionTokenLogprob {
+                    dynamo_protocols::types::ChatCompletionTokenLogprob {
                         token: t.clone(),
                         logprob: lp,
                         bytes: token_to_utf8_bytes(t),
@@ -226,7 +226,7 @@ impl DeltaGenerator {
                 .collect()
         });
 
-        Some(dynamo_async_openai::types::ChatChoiceLogprobs {
+        Some(dynamo_protocols::types::ChatChoiceLogprobs {
             content,
             refusal: None,
         })
@@ -242,22 +242,22 @@ impl DeltaGenerator {
     /// * `stop_reason` - Optional stop string or token that triggered the stop.
     ///
     /// # Returns
-    /// * An [`dynamo_async_openai::types::CreateChatCompletionStreamResponse`] instance representing the choice.
+    /// * An [`dynamo_protocols::types::CreateChatCompletionStreamResponse`] instance representing the choice.
     #[allow(deprecated)]
     pub fn create_choice(
         &mut self,
         index: u32,
         text: Option<String>,
-        finish_reason: Option<dynamo_async_openai::types::FinishReason>,
-        logprobs: Option<dynamo_async_openai::types::ChatChoiceLogprobs>,
-        stop_reason: Option<dynamo_async_openai::types::StopReason>,
+        finish_reason: Option<dynamo_protocols::types::FinishReason>,
+        logprobs: Option<dynamo_protocols::types::ChatChoiceLogprobs>,
+        stop_reason: Option<dynamo_protocols::types::StopReason>,
     ) -> NvCreateChatCompletionStreamResponse {
-        let delta = dynamo_async_openai::types::ChatCompletionStreamResponseDelta {
-            content: text.map(dynamo_async_openai::types::ChatCompletionMessageContent::Text),
+        let delta = dynamo_protocols::types::ChatCompletionStreamResponseDelta {
+            content: text.map(dynamo_protocols::types::ChatCompletionMessageContent::Text),
             function_call: None,
             tool_calls: None,
             role: if self.msg_counter == 0 {
-                Some(dynamo_async_openai::types::Role::Assistant)
+                Some(dynamo_protocols::types::Role::Assistant)
             } else {
                 None
             },
@@ -265,7 +265,7 @@ impl DeltaGenerator {
             reasoning_content: None,
         };
 
-        let choice = dynamo_async_openai::types::ChatChoiceStream {
+        let choice = dynamo_protocols::types::ChatChoiceStream {
             index,
             delta,
             finish_reason,
@@ -279,7 +279,7 @@ impl DeltaGenerator {
         // all intermediate chunks should have usage: null
         // The final usage chunk will be sent separately with empty choices
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: self.id.clone(),
                 object: self.object.clone(),
                 created: self.created,
@@ -306,7 +306,7 @@ impl DeltaGenerator {
         let usage = self.get_usage();
 
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: self.id.clone(),
                 object: self.object.clone(),
                 created: self.created,
@@ -330,7 +330,7 @@ impl DeltaGenerator {
         self.options.continuous_usage_stats
     }
 
-    pub fn get_usage(&self) -> dynamo_async_openai::types::CompletionUsage {
+    pub fn get_usage(&self) -> dynamo_protocols::types::CompletionUsage {
         let mut usage = self.usage.clone();
         usage.total_tokens = usage.prompt_tokens.saturating_add(usage.completion_tokens);
         usage
@@ -387,18 +387,16 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateChatCompletionStreamRes
 
         // Map backend finish reasons to OpenAI's finish reasons.
         let finish_reason = match delta.finish_reason {
-            Some(common::FinishReason::EoS) => Some(dynamo_async_openai::types::FinishReason::Stop),
-            Some(common::FinishReason::Stop) => {
-                Some(dynamo_async_openai::types::FinishReason::Stop)
-            }
+            Some(common::FinishReason::EoS) => Some(dynamo_protocols::types::FinishReason::Stop),
+            Some(common::FinishReason::Stop) => Some(dynamo_protocols::types::FinishReason::Stop),
             Some(common::FinishReason::Length) => {
-                Some(dynamo_async_openai::types::FinishReason::Length)
+                Some(dynamo_protocols::types::FinishReason::Length)
             }
             Some(common::FinishReason::Cancelled) => {
-                Some(dynamo_async_openai::types::FinishReason::Stop)
+                Some(dynamo_protocols::types::FinishReason::Stop)
             }
             Some(common::FinishReason::ContentFilter) => {
-                Some(dynamo_async_openai::types::FinishReason::ContentFilter)
+                Some(dynamo_protocols::types::FinishReason::ContentFilter)
             }
             Some(common::FinishReason::Error(err_msg)) => {
                 return Err(anyhow::anyhow!(err_msg));
@@ -490,7 +488,7 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateChatCompletionStreamRes
         DeltaGenerator::is_continuous_usage_enabled(self)
     }
 
-    fn get_usage(&self) -> dynamo_async_openai::types::CompletionUsage {
+    fn get_usage(&self) -> dynamo_protocols::types::CompletionUsage {
         DeltaGenerator::get_usage(self)
     }
 
@@ -502,7 +500,7 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateChatCompletionStreamRes
 #[cfg(test)]
 mod tests {
     use super::*;
-    use dynamo_async_openai::types::{
+    use dynamo_protocols::types::{
         ChatCompletionRequestMessage, ChatCompletionRequestUserMessage,
         ChatCompletionRequestUserMessageContent, CreateChatCompletionRequest,
     };

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

@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 
 use async_stream::stream;
-use dynamo_async_openai::types::{
+use dynamo_protocols::types::{
     ChatChoiceLogprobs, ChatChoiceStream, ChatCompletionMessageToolCallChunk,
     ChatCompletionStreamResponseDelta, FinishReason, FunctionCallStream, Role,
 };
@@ -116,7 +116,7 @@ fn create_choice_stream(
     content: &str,
     tool_calls: Option<Vec<ChatCompletionMessageToolCallChunk>>,
     finish_reason: Option<FinishReason>,
-    stop_reason: Option<dynamo_async_openai::types::StopReason>,
+    stop_reason: Option<dynamo_protocols::types::StopReason>,
     logprobs: Option<ChatChoiceLogprobs>,
 ) -> ChatChoiceStream {
     #[allow(deprecated)]
@@ -124,9 +124,9 @@ fn create_choice_stream(
         index,
         delta: ChatCompletionStreamResponseDelta {
             role,
-            content: Some(
-                dynamo_async_openai::types::ChatCompletionMessageContent::Text(content.to_string()),
-            ),
+            content: Some(dynamo_protocols::types::ChatCompletionMessageContent::Text(
+                content.to_string(),
+            )),
             tool_calls,
             function_call: None,
             refusal: None,
@@ -543,8 +543,8 @@ impl JailedStream {
                         if let Some(ref content) = choice.delta.content {
                             // Jailing only applies to text content
                             let text_content = match content {
-                                dynamo_async_openai::types::ChatCompletionMessageContent::Text(text) => Some(text.as_str()),
-                                dynamo_async_openai::types::ChatCompletionMessageContent::Parts(_) => None,
+                                dynamo_protocols::types::ChatCompletionMessageContent::Text(text) => Some(text.as_str()),
+                                dynamo_protocols::types::ChatCompletionMessageContent::Parts(_) => None,
                             };
 
                             if let Some(text) = text_content {
@@ -676,7 +676,7 @@ impl JailedStream {
                 tracing::debug!("Stream ended while jailed, releasing accumulated content");
                 // Create a finalization response carrying forward real stream metadata
                 let dummy_response = NvCreateChatCompletionStreamResponse {
-                    inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+                    inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                         id: last_stream_id,
                         object: "chat.completion.chunk".to_string(),
                         created: last_stream_created,
@@ -932,7 +932,7 @@ impl JailedStream {
         ChatCompletionMessageToolCallChunk {
             index,
             id: Some(format!("call-{}", Uuid::new_v4())),
-            r#type: Some(dynamo_async_openai::types::ChatCompletionToolType::Function),
+            r#type: Some(dynamo_protocols::types::ChatCompletionToolType::Function),
             function: Some(FunctionCallStream {
                 name: Some(name),
                 arguments: Some(arguments),

lib/llm/src/protocols/openai/completions.rs

@@ -27,7 +27,7 @@ pub use delta::DeltaGenerator;
 #[derive(ToSchema, Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvCreateCompletionRequest {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateCompletionRequest,
+    pub inner: dynamo_protocols::types::CreateCompletionRequest,
 
     #[serde(flatten)]
     pub common: CommonExt,
@@ -47,27 +47,27 @@ pub struct NvCreateCompletionRequest {
 #[derive(ToSchema, Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvCreateCompletionResponse {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateCompletionResponse,
+    pub inner: dynamo_protocols::types::CreateCompletionResponse,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub nvext: Option<serde_json::Value>,
 }
 
-impl ContentProvider for dynamo_async_openai::types::Choice {
+impl ContentProvider for dynamo_protocols::types::Choice {
     fn content(&self) -> String {
         self.text.clone()
     }
 }
 
-pub fn prompt_to_string(prompt: &dynamo_async_openai::types::Prompt) -> String {
+pub fn prompt_to_string(prompt: &dynamo_protocols::types::Prompt) -> String {
     match prompt {
-        dynamo_async_openai::types::Prompt::String(s) => s.clone(),
-        dynamo_async_openai::types::Prompt::StringArray(arr) => arr.join(" "), // Join strings with spaces
-        dynamo_async_openai::types::Prompt::IntegerArray(arr) => arr
+        dynamo_protocols::types::Prompt::String(s) => s.clone(),
+        dynamo_protocols::types::Prompt::StringArray(arr) => arr.join(" "), // Join strings with spaces
+        dynamo_protocols::types::Prompt::IntegerArray(arr) => arr
             .iter()
             .map(|&num| num.to_string())
             .collect::<Vec<_>>()
             .join(" "),
-        dynamo_async_openai::types::Prompt::ArrayOfIntegerArray(arr) => arr
+        dynamo_protocols::types::Prompt::ArrayOfIntegerArray(arr) => arr
             .iter()
             .map(|inner| {
                 inner
@@ -82,12 +82,12 @@ pub fn prompt_to_string(prompt: &dynamo_async_openai::types::Prompt) -> String {
 }
 
 /// Get the batch size from a prompt (1 for single prompts, array length for batch prompts)
-pub fn get_prompt_batch_size(prompt: &dynamo_async_openai::types::Prompt) -> usize {
+pub fn get_prompt_batch_size(prompt: &dynamo_protocols::types::Prompt) -> usize {
     match prompt {
-        dynamo_async_openai::types::Prompt::String(_) => 1,
-        dynamo_async_openai::types::Prompt::IntegerArray(_) => 1,
-        dynamo_async_openai::types::Prompt::StringArray(arr) => arr.len(),
-        dynamo_async_openai::types::Prompt::ArrayOfIntegerArray(arr) => arr.len(),
+        dynamo_protocols::types::Prompt::String(_) => 1,
+        dynamo_protocols::types::Prompt::IntegerArray(_) => 1,
+        dynamo_protocols::types::Prompt::StringArray(arr) => arr.len(),
+        dynamo_protocols::types::Prompt::ArrayOfIntegerArray(arr) => arr.len(),
     }
 }
 
@@ -95,21 +95,21 @@ pub fn get_prompt_batch_size(prompt: &dynamo_async_openai::types::Prompt) -> usi
 /// For single prompts, returns a clone regardless of index.
 /// For batch prompts, returns the prompt at the specified index.
 pub fn extract_single_prompt(
-    prompt: &dynamo_async_openai::types::Prompt,
+    prompt: &dynamo_protocols::types::Prompt,
     index: usize,
-) -> dynamo_async_openai::types::Prompt {
+) -> dynamo_protocols::types::Prompt {
     match prompt {
-        dynamo_async_openai::types::Prompt::String(s) => {
-            dynamo_async_openai::types::Prompt::String(s.clone())
+        dynamo_protocols::types::Prompt::String(s) => {
+            dynamo_protocols::types::Prompt::String(s.clone())
         }
-        dynamo_async_openai::types::Prompt::IntegerArray(arr) => {
-            dynamo_async_openai::types::Prompt::IntegerArray(arr.clone())
+        dynamo_protocols::types::Prompt::IntegerArray(arr) => {
+            dynamo_protocols::types::Prompt::IntegerArray(arr.clone())
         }
-        dynamo_async_openai::types::Prompt::StringArray(arr) => {
-            dynamo_async_openai::types::Prompt::String(arr[index].clone())
+        dynamo_protocols::types::Prompt::StringArray(arr) => {
+            dynamo_protocols::types::Prompt::String(arr[index].clone())
         }
-        dynamo_async_openai::types::Prompt::ArrayOfIntegerArray(arr) => {
-            dynamo_async_openai::types::Prompt::IntegerArray(arr[index].clone())
+        dynamo_protocols::types::Prompt::ArrayOfIntegerArray(arr) => {
+            dynamo_protocols::types::Prompt::IntegerArray(arr[index].clone())
         }
     }
 }
@@ -241,7 +241,7 @@ impl OpenAIStopConditionsProvider for NvCreateCompletionRequest {
     }
 
     fn get_stop(&self) -> Option<Vec<String>> {
-        use dynamo_async_openai::types::Stop;
+        use dynamo_protocols::types::Stop;
         self.inner.stop.as_ref().map(|s| match s {
             Stop::String(s) => vec![s.clone()],
             Stop::StringArray(arr) => arr.clone(),
@@ -287,10 +287,10 @@ impl ResponseFactory {
 
     pub fn make_response(
         &self,
-        choice: dynamo_async_openai::types::Choice,
-        usage: Option<dynamo_async_openai::types::CompletionUsage>,
+        choice: dynamo_protocols::types::Choice,
+        usage: Option<dynamo_protocols::types::CompletionUsage>,
     ) -> NvCreateCompletionResponse {
-        let inner = dynamo_async_openai::types::CreateCompletionResponse {
+        let inner = dynamo_protocols::types::CreateCompletionResponse {
             id: self.id.clone(),
             object: self.object.clone(),
             created: self.created,
@@ -361,7 +361,7 @@ impl TryFrom<NvCreateCompletionRequest> for common::CompletionRequest {
     }
 }
 
-impl TryFrom<common::StreamingCompletionResponse> for dynamo_async_openai::types::Choice {
+impl TryFrom<common::StreamingCompletionResponse> for dynamo_protocols::types::Choice {
     type Error = anyhow::Error;
 
     fn try_from(response: common::StreamingCompletionResponse) -> Result<Self, Self::Error> {
@@ -382,10 +382,10 @@ impl TryFrom<common::StreamingCompletionResponse> for dynamo_async_openai::types
         // TODO handle aggregating logprobs
         let logprobs = None;
 
-        let finish_reason: Option<dynamo_async_openai::types::CompletionFinishReason> =
+        let finish_reason: Option<dynamo_protocols::types::CompletionFinishReason> =
             response.delta.finish_reason.map(Into::into);
 
-        let choice = dynamo_async_openai::types::Choice {
+        let choice = dynamo_protocols::types::Choice {
             text,
             index,
             logprobs,

lib/llm/src/protocols/openai/completions/aggregator.rs

@@ -20,7 +20,7 @@ pub struct DeltaAggregator {
     id: String,
     model: String,
     created: u32,
-    usage: Option<dynamo_async_openai::types::CompletionUsage>,
+    usage: Option<dynamo_protocols::types::CompletionUsage>,
     system_fingerprint: Option<String>,
     choices: HashMap<u32, DeltaChoice>,
     error: Option<String>,
@@ -31,7 +31,7 @@ struct DeltaChoice {
     index: u32,
     text: String,
     finish_reason: Option<FinishReason>,
-    logprobs: Option<dynamo_async_openai::types::Logprobs>,
+    logprobs: Option<dynamo_protocols::types::Logprobs>,
 }
 
 impl Default for DeltaAggregator {
@@ -109,14 +109,14 @@ impl DeltaAggregator {
 
                         // Handle CompletionFinishReason -> FinishReason conversation
                         state_choice.finish_reason = match choice.finish_reason {
-                            Some(dynamo_async_openai::types::CompletionFinishReason::Stop) => {
+                            Some(dynamo_protocols::types::CompletionFinishReason::Stop) => {
                                 Some(FinishReason::Stop)
                             }
-                            Some(dynamo_async_openai::types::CompletionFinishReason::Length) => {
+                            Some(dynamo_protocols::types::CompletionFinishReason::Length) => {
                                 Some(FinishReason::Length)
                             }
                             Some(
-                                dynamo_async_openai::types::CompletionFinishReason::ContentFilter,
+                                dynamo_protocols::types::CompletionFinishReason::ContentFilter,
                             ) => Some(FinishReason::ContentFilter),
                             None => None,
                         };
@@ -124,7 +124,7 @@ impl DeltaAggregator {
                         // Update logprobs
                         if let Some(logprobs) = &choice.logprobs {
                             let state_lps = state_choice.logprobs.get_or_insert(
-                                dynamo_async_openai::types::Logprobs {
+                                dynamo_protocols::types::Logprobs {
                                     tokens: Vec::new(),
                                     token_logprobs: Vec::new(),
                                     top_logprobs: Vec::new(),
@@ -155,12 +155,12 @@ impl DeltaAggregator {
         let mut choices: Vec<_> = aggregator
             .choices
             .into_values()
-            .map(dynamo_async_openai::types::Choice::from)
+            .map(dynamo_protocols::types::Choice::from)
             .collect();
 
         choices.sort_by(|a, b| a.index.cmp(&b.index));
 
-        let inner = dynamo_async_openai::types::CreateCompletionResponse {
+        let inner = dynamo_protocols::types::CreateCompletionResponse {
             id: aggregator.id,
             created: aggregator.created,
             usage: aggregator.usage,
@@ -179,11 +179,11 @@ impl DeltaAggregator {
     }
 }
 
-impl From<DeltaChoice> for dynamo_async_openai::types::Choice {
+impl From<DeltaChoice> for dynamo_protocols::types::Choice {
     fn from(delta: DeltaChoice) -> Self {
         let finish_reason = delta.finish_reason.map(Into::into);
 
-        dynamo_async_openai::types::Choice {
+        dynamo_protocols::types::Choice {
             index: delta.index,
             text: delta.text,
             finish_reason,
@@ -231,11 +231,11 @@ mod tests {
             .and_then(|s| FinishReason::from_str(s).ok())
             .map(Into::into);
 
-        let logprobs = logprob.map(|lp| dynamo_async_openai::types::Logprobs {
+        let logprobs = logprob.map(|lp| dynamo_protocols::types::Logprobs {
             tokens: vec![text.to_string()],
             token_logprobs: vec![Some(lp)],
             top_logprobs: vec![
-                serde_json::to_value(dynamo_async_openai::types::TopLogprobs {
+                serde_json::to_value(dynamo_protocols::types::TopLogprobs {
                     token: text.to_string(),
                     logprob: lp,
                     bytes: None,
@@ -245,13 +245,13 @@ mod tests {
             text_offset: vec![0],
         });
 
-        let inner = dynamo_async_openai::types::CreateCompletionResponse {
+        let inner = dynamo_protocols::types::CreateCompletionResponse {
             id: "test_id".to_string(),
             model: "meta/llama-3.1-8b".to_string(),
             created: 1234567890,
             usage: None,
             system_fingerprint: None,
-            choices: vec![dynamo_async_openai::types::Choice {
+            choices: vec![dynamo_protocols::types::Choice {
                 index,
                 text: text.to_string(),
                 finish_reason,
@@ -319,11 +319,11 @@ mod tests {
         assert_eq!(choice.text, "Hello,".to_string());
         assert_eq!(
             choice.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Length)
+            Some(dynamo_protocols::types::CompletionFinishReason::Length)
         );
         assert_eq!(
             choice.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Length)
+            Some(dynamo_protocols::types::CompletionFinishReason::Length)
         );
         assert!(choice.logprobs.is_none());
     }
@@ -355,7 +355,7 @@ mod tests {
         assert_eq!(choice.text, "Hello, world!".to_string());
         assert_eq!(
             choice.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Stop)
+            Some(dynamo_protocols::types::CompletionFinishReason::Stop)
         );
         assert_eq!(choice.logprobs.as_ref().unwrap().tokens.len(), 2);
         assert_eq!(
@@ -367,23 +367,23 @@ mod tests {
     #[tokio::test]
     async fn test_multiple_choices() {
         // Create a delta with multiple choices
-        let inner = dynamo_async_openai::types::CreateCompletionResponse {
+        let inner = dynamo_protocols::types::CreateCompletionResponse {
             id: "test_id".to_string(),
             model: "meta/llama-3.1-8b".to_string(),
             created: 1234567890,
             usage: None,
             system_fingerprint: None,
             choices: vec![
-                dynamo_async_openai::types::Choice {
+                dynamo_protocols::types::Choice {
                     index: 0,
                     text: "Choice 0".to_string(),
-                    finish_reason: Some(dynamo_async_openai::types::CompletionFinishReason::Stop),
+                    finish_reason: Some(dynamo_protocols::types::CompletionFinishReason::Stop),
                     logprobs: None,
                 },
-                dynamo_async_openai::types::Choice {
+                dynamo_protocols::types::Choice {
                     index: 1,
                     text: "Choice 1".to_string(),
-                    finish_reason: Some(dynamo_async_openai::types::CompletionFinishReason::Stop),
+                    finish_reason: Some(dynamo_protocols::types::CompletionFinishReason::Stop),
                     logprobs: None,
                 },
             ],
@@ -418,11 +418,11 @@ mod tests {
         assert_eq!(choice0.text, "Choice 0".to_string());
         assert_eq!(
             choice0.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Stop)
+            Some(dynamo_protocols::types::CompletionFinishReason::Stop)
         );
         assert_eq!(
             choice0.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Stop)
+            Some(dynamo_protocols::types::CompletionFinishReason::Stop)
         );
 
         let choice1 = &response.inner.choices[1];
@@ -430,11 +430,11 @@ mod tests {
         assert_eq!(choice1.text, "Choice 1".to_string());
         assert_eq!(
             choice1.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Stop)
+            Some(dynamo_protocols::types::CompletionFinishReason::Stop)
         );
         assert_eq!(
             choice1.finish_reason,
-            Some(dynamo_async_openai::types::CompletionFinishReason::Stop)
+            Some(dynamo_protocols::types::CompletionFinishReason::Stop)
         );
     }
 }

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

@@ -31,7 +31,7 @@ impl NvCreateCompletionRequest {
             // For non-streaming requests (stream=false), enable usage by default
             if self.inner.stream_options.is_none() {
                 self.inner.stream_options =
-                    Some(dynamo_async_openai::types::ChatCompletionStreamOptions {
+                    Some(dynamo_protocols::types::ChatCompletionStreamOptions {
                         include_usage: true,
                         continuous_usage_stats: false,
                     });
@@ -95,7 +95,7 @@ pub struct DeltaGenerator {
     created: u32,
     model: String,
     system_fingerprint: Option<String>,
-    usage: dynamo_async_openai::types::CompletionUsage,
+    usage: dynamo_protocols::types::CompletionUsage,
     options: DeltaGeneratorOptions,
     tracker: Option<Arc<RequestTracker>>,
 }
@@ -113,7 +113,7 @@ impl DeltaGenerator {
 
         // Previously, our home-rolled CompletionUsage impl'd Default
         // PR !387 - https://github.com/64bit/async-openai/pull/387
-        let usage = dynamo_async_openai::types::CompletionUsage {
+        let usage = dynamo_protocols::types::CompletionUsage {
             completion_tokens: 0,
             prompt_tokens: 0,
             total_tokens: 0,
@@ -154,7 +154,7 @@ impl DeltaGenerator {
         token_ids: Vec<TokenIdType>,
         logprobs: Option<common::llm_backend::LogProbs>,
         top_logprobs: Option<common::llm_backend::TopLogprobs>,
-    ) -> Option<dynamo_async_openai::types::Logprobs> {
+    ) -> Option<dynamo_protocols::types::Logprobs> {
         if !self.options.enable_logprobs || logprobs.is_none() {
             return None;
         }
@@ -181,7 +181,7 @@ impl DeltaGenerator {
                 .collect()
         });
 
-        Some(dynamo_async_openai::types::Logprobs {
+        Some(dynamo_protocols::types::Logprobs {
             tokens: toks.iter().map(|(t, _)| t.clone()).collect(),
             token_logprobs: tok_lps.into_iter().map(Some).collect(),
             text_offset: vec![],
@@ -193,21 +193,21 @@ impl DeltaGenerator {
         &self,
         index: u32,
         text: Option<String>,
-        finish_reason: Option<dynamo_async_openai::types::CompletionFinishReason>,
-        logprobs: Option<dynamo_async_openai::types::Logprobs>,
+        finish_reason: Option<dynamo_protocols::types::CompletionFinishReason>,
+        logprobs: Option<dynamo_protocols::types::Logprobs>,
     ) -> NvCreateCompletionResponse {
         // todo - update for tool calling
 
         // According to OpenAI spec: when stream_options.include_usage is true,
         // all intermediate chunks should have usage: null
         // The final usage chunk will be sent separately with empty choices
-        let inner = dynamo_async_openai::types::CreateCompletionResponse {
+        let inner = dynamo_protocols::types::CreateCompletionResponse {
             id: self.id.clone(),
             object: self.object.clone(),
             created: self.created,
             model: self.model.clone(),
             system_fingerprint: self.system_fingerprint.clone(),
-            choices: vec![dynamo_async_openai::types::Choice {
+            choices: vec![dynamo_protocols::types::Choice {
                 text: text.unwrap_or_default(),
                 index,
                 finish_reason,
@@ -231,7 +231,7 @@ impl DeltaGenerator {
     pub fn create_usage_chunk(&self) -> NvCreateCompletionResponse {
         let usage = self.get_usage();
 
-        let inner = dynamo_async_openai::types::CreateCompletionResponse {
+        let inner = dynamo_protocols::types::CreateCompletionResponse {
             id: self.id.clone(),
             object: self.object.clone(),
             created: self.created,
@@ -254,7 +254,7 @@ impl DeltaGenerator {
         self.options.continuous_usage_stats
     }
 
-    pub fn get_usage(&self) -> dynamo_async_openai::types::CompletionUsage {
+    pub fn get_usage(&self) -> dynamo_protocols::types::CompletionUsage {
         let mut usage = self.usage.clone();
         usage.total_tokens = usage.prompt_tokens.saturating_add(usage.completion_tokens);
         usage
@@ -377,7 +377,7 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateCompletionResponse> for
         DeltaGenerator::is_continuous_usage_enabled(self)
     }
 
-    fn get_usage(&self) -> dynamo_async_openai::types::CompletionUsage {
+    fn get_usage(&self) -> dynamo_protocols::types::CompletionUsage {
         DeltaGenerator::get_usage(self)
     }
 

lib/llm/src/protocols/openai/embeddings.rs

@@ -15,7 +15,7 @@ pub use nvext::{NvExt, NvExtProvider};
 #[derive(ToSchema, Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvCreateEmbeddingRequest {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateEmbeddingRequest,
+    pub inner: dynamo_protocols::types::CreateEmbeddingRequest,
 
     #[serde(skip_serializing_if = "Option::is_none")]
     pub nvext: Option<NvExt>,
@@ -30,17 +30,17 @@ pub struct NvCreateEmbeddingRequest {
 #[derive(ToSchema, Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvCreateEmbeddingResponse {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateEmbeddingResponse,
+    pub inner: dynamo_protocols::types::CreateEmbeddingResponse,
 }
 
 impl NvCreateEmbeddingResponse {
     pub fn empty() -> Self {
         Self {
-            inner: dynamo_async_openai::types::CreateEmbeddingResponse {
+            inner: dynamo_protocols::types::CreateEmbeddingResponse {
                 object: "list".to_string(),
                 model: "embedding".to_string(),
                 data: vec![],
-                usage: dynamo_async_openai::types::EmbeddingUsage {
+                usage: dynamo_protocols::types::EmbeddingUsage {
                     prompt_tokens: 0,
                     total_tokens: 0,
                 },

lib/llm/src/protocols/openai/embeddings/aggregator.rs

@@ -133,16 +133,16 @@ mod tests {
     use futures::stream;
 
     fn create_test_embedding_response(
-        embeddings: Vec<dynamo_async_openai::types::Embedding>,
+        embeddings: Vec<dynamo_protocols::types::Embedding>,
         prompt_tokens: u32,
         total_tokens: u32,
     ) -> Annotated<NvCreateEmbeddingResponse> {
         let response = NvCreateEmbeddingResponse {
-            inner: dynamo_async_openai::types::CreateEmbeddingResponse {
+            inner: dynamo_protocols::types::CreateEmbeddingResponse {
                 object: "list".to_string(),
                 model: "test-model".to_string(),
                 data: embeddings,
-                usage: dynamo_async_openai::types::EmbeddingUsage {
+                usage: dynamo_protocols::types::EmbeddingUsage {
                     prompt_tokens,
                     total_tokens,
                 },
@@ -166,7 +166,7 @@ mod tests {
 
     #[tokio::test]
     async fn test_single_embedding() {
-        let embedding = dynamo_async_openai::types::Embedding {
+        let embedding = dynamo_protocols::types::Embedding {
             index: 0,
             object: "embedding".to_string(),
             embedding: vec![0.1, 0.2, 0.3],
@@ -188,13 +188,13 @@ mod tests {
 
     #[tokio::test]
     async fn test_multiple_embeddings() {
-        let embedding1 = dynamo_async_openai::types::Embedding {
+        let embedding1 = dynamo_protocols::types::Embedding {
             index: 0,
             object: "embedding".to_string(),
             embedding: vec![0.1, 0.2, 0.3],
         };
 
-        let embedding2 = dynamo_async_openai::types::Embedding {
+        let embedding2 = dynamo_protocols::types::Embedding {
             index: 1,
             object: "embedding".to_string(),
             embedding: vec![0.4, 0.5, 0.6],

lib/llm/src/protocols/openai/images.rs

@@ -14,7 +14,7 @@ pub use nvext::{NvExt, NvExtProvider};
 #[derive(Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvCreateImageRequest {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::CreateImageRequest,
+    pub inner: dynamo_protocols::types::CreateImageRequest,
 
     #[serde(skip_serializing_if = "Option::is_none")]
     pub nvext: Option<NvExt>,
@@ -28,13 +28,13 @@ pub struct NvCreateImageRequest {
 #[derive(Serialize, Deserialize, Validate, Debug, Clone)]
 pub struct NvImagesResponse {
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::ImagesResponse,
+    pub inner: dynamo_protocols::types::ImagesResponse,
 }
 
 impl NvImagesResponse {
     pub fn empty() -> Self {
         Self {
-            inner: dynamo_async_openai::types::ImagesResponse {
+            inner: dynamo_protocols::types::ImagesResponse {
                 created: 0,
                 data: vec![],
             },

lib/llm/src/protocols/openai/nvext.rs

@@ -216,7 +216,7 @@ pub struct AgentHints {
 
 // Re-export CacheControl types from dynamo-async-openai where they are canonically defined
 // alongside the Anthropic protocol types they originate from.
-pub use dynamo_async_openai::types::anthropic::{CacheControl, CacheControlType};
+pub use dynamo_protocols::types::anthropic::{CacheControl, CacheControlType};
 
 impl Default for NvExt {
     fn default() -> Self {

lib/llm/src/protocols/openai/responses/mod.rs

@@ -3,7 +3,7 @@
 
 pub mod stream_converter;
 
-use dynamo_async_openai::types::responses::{
+use dynamo_protocols::types::responses::{
     AssistantRole, FunctionCallOutput, FunctionToolCall, IncludeEnum, InputContent, InputItem,
     InputParam, InputRole, InputTokenDetails, Instructions, Item, MessageItem, OutputItem,
     OutputMessage, OutputMessageContent, OutputStatus, OutputTextContent, OutputTokenDetails,
@@ -11,7 +11,7 @@ use dynamo_async_openai::types::responses::{
     ServiceTier, Status, Summary, SummaryPart, TextResponseFormatConfiguration, Tool,
     ToolChoiceOptions, ToolChoiceParam, Truncation,
 };
-use dynamo_async_openai::types::{
+use dynamo_protocols::types::{
     ChatCompletionMessageToolCall, ChatCompletionNamedToolChoice,
     ChatCompletionRequestAssistantMessage, ChatCompletionRequestAssistantMessageContent,
     ChatCompletionRequestMessage, ChatCompletionRequestMessageContentPartImage,
@@ -38,7 +38,7 @@ use super::{OpenAISamplingOptionsProvider, OpenAIStopConditionsProvider};
 pub struct NvCreateResponse {
     /// Flattened CreateResponse fields (model, input, temperature, etc.)
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::responses::CreateResponse,
+    pub inner: dynamo_protocols::types::responses::CreateResponse,
 
     #[serde(skip_serializing_if = "Option::is_none")]
     pub nvext: Option<NvExt>,
@@ -48,7 +48,7 @@ pub struct NvCreateResponse {
 pub struct NvResponse {
     /// Flattened Response fields.
     #[serde(flatten)]
-    pub inner: dynamo_async_openai::types::responses::Response,
+    pub inner: dynamo_protocols::types::responses::Response,
 
     /// NVIDIA extension field for response metadata (worker IDs, etc.)
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -144,12 +144,12 @@ impl OpenAIStopConditionsProvider for NvCreateResponse {
 
 /// Convert a Responses API ImageDetail to the Chat Completions ImageDetail.
 fn convert_image_detail(
-    detail: &dynamo_async_openai::types::responses::ImageDetail,
+    detail: &dynamo_protocols::types::responses::ImageDetail,
 ) -> ChatImageDetail {
     match detail {
-        dynamo_async_openai::types::responses::ImageDetail::Auto => ChatImageDetail::Auto,
-        dynamo_async_openai::types::responses::ImageDetail::Low => ChatImageDetail::Low,
-        dynamo_async_openai::types::responses::ImageDetail::High => ChatImageDetail::High,
+        dynamo_protocols::types::responses::ImageDetail::Auto => ChatImageDetail::Auto,
+        dynamo_protocols::types::responses::ImageDetail::Low => ChatImageDetail::Low,
+        dynamo_protocols::types::responses::ImageDetail::High => ChatImageDetail::High,
     }
 }
 
@@ -316,7 +316,7 @@ fn convert_input_items_to_messages(
                             tool_calls: Some(vec![ChatCompletionMessageToolCall {
                                 id: fc.call_id.clone(),
                                 r#type: ChatCompletionToolType::Function,
-                                function: dynamo_async_openai::types::FunctionCall {
+                                function: dynamo_protocols::types::FunctionCall {
                                     name: fc.name.clone(),
                                     arguments: fc.arguments.clone(),
                                 },
@@ -349,10 +349,10 @@ fn convert_input_items_to_messages(
             InputItem::EasyMessage(easy) => {
                 // Handle easy input messages based on role
                 let content_text = match &easy.content {
-                    dynamo_async_openai::types::responses::EasyInputContent::Text(text) => {
+                    dynamo_protocols::types::responses::EasyInputContent::Text(text) => {
                         text.clone()
                     }
-                    dynamo_async_openai::types::responses::EasyInputContent::ContentList(parts) => {
+                    dynamo_protocols::types::responses::EasyInputContent::ContentList(parts) => {
                         convert_input_content_to_text(parts)
                     }
                 };
@@ -737,10 +737,8 @@ pub fn chat_completion_to_response(
         // Handle text content -- also parse <tool_call> blocks from models
         // that emit tool calls as text (e.g. Qwen3)
         let content_text = match choice.message.content {
-            Some(dynamo_async_openai::types::ChatCompletionMessageContent::Text(text)) => {
-                Some(text)
-            }
-            Some(dynamo_async_openai::types::ChatCompletionMessageContent::Parts(_)) => {
+            Some(dynamo_protocols::types::ChatCompletionMessageContent::Text(text)) => Some(text),
+            Some(dynamo_protocols::types::ChatCompletionMessageContent::Parts(_)) => {
                 tracing::warn!(
                     "Multimodal content in responses API not yet supported, using placeholder"
                 );
@@ -880,12 +878,12 @@ pub fn chat_completion_to_response(
 
 #[cfg(test)]
 mod tests {
-    use dynamo_async_openai::types::responses::{
+    use dynamo_protocols::types::responses::{
         CreateResponse, FunctionCallOutput, FunctionCallOutputItemParam, FunctionTool,
         FunctionToolCall, ImageDetail, InputContent, InputImageContent, InputItem, InputMessage,
         InputParam, InputRole, InputTextContent, Item, MessageItem, Tool,
     };
-    use dynamo_async_openai::types::{
+    use dynamo_protocols::types::{
         ChatCompletionRequestMessage, ChatCompletionRequestUserMessageContent,
     };
 
@@ -1167,19 +1165,17 @@ mod tests {
     fn test_into_nvresponse_from_chat_response() {
         let now = 1_726_000_000;
         let chat_resp = NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 id: "chatcmpl-xyz".into(),
-                choices: vec![dynamo_async_openai::types::ChatChoice {
+                choices: vec![dynamo_protocols::types::ChatChoice {
                     index: 0,
-                    message: dynamo_async_openai::types::ChatCompletionResponseMessage {
-                        content: Some(
-                            dynamo_async_openai::types::ChatCompletionMessageContent::Text(
+                    message: dynamo_protocols::types::ChatCompletionResponseMessage {
+                        content: Some(dynamo_protocols::types::ChatCompletionMessageContent::Text(
                             "This is a reply".to_string(),
-                            ),
-                        ),
+                        )),
                         refusal: None,
                         tool_calls: None,
-                        role: dynamo_async_openai::types::Role::Assistant,
+                        role: dynamo_protocols::types::Role::Assistant,
                         function_call: None,
                         audio: None,
                         reasoning_content: None,
@@ -1225,22 +1221,22 @@ mod tests {
     fn test_response_with_tool_calls() {
         let now = 1_726_000_000;
         let chat_resp = NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 id: "chatcmpl-xyz".into(),
-                choices: vec![dynamo_async_openai::types::ChatChoice {
+                choices: vec![dynamo_protocols::types::ChatChoice {
                     index: 0,
-                    message: dynamo_async_openai::types::ChatCompletionResponseMessage {
+                    message: dynamo_protocols::types::ChatCompletionResponseMessage {
                         content: None,
                         refusal: None,
                         tool_calls: Some(vec![ChatCompletionMessageToolCall {
                             id: "call_abc".into(),
                             r#type: ChatCompletionToolType::Function,
-                            function: dynamo_async_openai::types::FunctionCall {
+                            function: dynamo_protocols::types::FunctionCall {
                                 name: "get_weather".into(),
                                 arguments: r#"{"location":"SF"}"#.into(),
                             },
                         }]),
-                        role: dynamo_async_openai::types::Role::Assistant,
+                        role: dynamo_protocols::types::Role::Assistant,
                         function_call: None,
                         audio: None,
                         reasoning_content: None,
@@ -1335,8 +1331,8 @@ thinking
 
     #[test]
     fn test_reasoning_effort_mapped_to_chat_completion() {
-        use dynamo_async_openai::types::ReasoningEffort;
-        use dynamo_async_openai::types::responses::Reasoning;
+        use dynamo_protocols::types::ReasoningEffort;
+        use dynamo_protocols::types::responses::Reasoning;
 
         let mut req = make_response_with_input("think hard");
         req.inner.reasoning = Some(Reasoning {
@@ -1357,8 +1353,8 @@ thinking
 
     #[test]
     fn test_text_format_json_object_mapped() {
-        use dynamo_async_openai::types::ResponseFormat;
-        use dynamo_async_openai::types::responses::{
+        use dynamo_protocols::types::ResponseFormat;
+        use dynamo_protocols::types::responses::{
             ResponseTextParam, TextResponseFormatConfiguration,
         };
 
@@ -1374,10 +1370,10 @@ thinking
 
     #[test]
     fn test_text_format_json_schema_mapped() {
-        use dynamo_async_openai::types::responses::{
+        use dynamo_protocols::types::responses::{
             ResponseTextParam, TextResponseFormatConfiguration,
         };
-        use dynamo_async_openai::types::{ResponseFormat, ResponseFormatJsonSchema};
+        use dynamo_protocols::types::{ResponseFormat, ResponseFormatJsonSchema};
 
         let schema = ResponseFormatJsonSchema {
             name: "city".into(),
@@ -1402,7 +1398,7 @@ thinking
 
     #[test]
     fn test_text_format_plain_text_leaves_response_format_none() {
-        use dynamo_async_openai::types::responses::{
+        use dynamo_protocols::types::responses::{
             ResponseTextParam, TextResponseFormatConfiguration,
         };
 
@@ -1418,8 +1414,8 @@ thinking
 
     #[test]
     fn test_service_tier_mapped_to_chat_completion() {
-        use dynamo_async_openai::types::ServiceTier as ChatServiceTier;
-        use dynamo_async_openai::types::responses::ServiceTier as RespServiceTier;
+        use dynamo_protocols::types::ServiceTier as ChatServiceTier;
+        use dynamo_protocols::types::responses::ServiceTier as RespServiceTier;
 
         let mut req = make_response_with_input("priority");
         req.inner.service_tier = Some(RespServiceTier::Priority);
@@ -1430,8 +1426,8 @@ thinking
 
     #[test]
     fn test_response_echoes_reasoning() {
-        use dynamo_async_openai::types::ReasoningEffort;
-        use dynamo_async_openai::types::responses::Reasoning;
+        use dynamo_protocols::types::ReasoningEffort;
+        use dynamo_protocols::types::responses::Reasoning;
 
         let params = ResponseParams {
             reasoning: Some(Reasoning {
@@ -1442,7 +1438,7 @@ thinking
         };
 
         let chat_resp = NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 choices: vec![],
                 created: 0,
                 id: "test".into(),
@@ -1462,7 +1458,7 @@ thinking
 
     #[test]
     fn test_response_echoes_text_format() {
-        use dynamo_async_openai::types::responses::{
+        use dynamo_protocols::types::responses::{
             ResponseTextParam, TextResponseFormatConfiguration,
         };
 
@@ -1475,7 +1471,7 @@ thinking
         };
 
         let chat_resp = NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 choices: vec![],
                 created: 0,
                 id: "test".into(),
@@ -1495,7 +1491,7 @@ thinking
 
     #[test]
     fn test_response_echoes_service_tier() {
-        use dynamo_async_openai::types::responses::ServiceTier;
+        use dynamo_protocols::types::responses::ServiceTier;
 
         let params = ResponseParams {
             service_tier: Some(ServiceTier::Flex),
@@ -1503,7 +1499,7 @@ thinking
         };
 
         let chat_resp = NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 choices: vec![],
                 created: 0,
                 id: "test".into(),
@@ -1522,7 +1518,7 @@ thinking
 
     #[test]
     fn test_output_message_deserializes_without_id_and_status() {
-        use dynamo_async_openai::types::responses::{InputItem, Item, MessageItem};
+        use dynamo_protocols::types::responses::{InputItem, Item, MessageItem};
 
         let json = serde_json::json!({
             "role": "assistant",
@@ -1544,7 +1540,7 @@ thinking
 
     #[test]
     fn test_output_message_with_id_and_status_still_works() {
-        use dynamo_async_openai::types::responses::{InputItem, Item, MessageItem, OutputStatus};
+        use dynamo_protocols::types::responses::{InputItem, Item, MessageItem, OutputStatus};
 
         let json = serde_json::json!({
             "role": "assistant",
@@ -1567,17 +1563,17 @@ thinking
     // ── PR2: include filtering + truncation echo-back tests ──
 
     fn make_chat_resp_with_text(text: &str) -> NvCreateChatCompletionResponse {
-        use dynamo_async_openai::types::{
+        use dynamo_protocols::types::{
             ChatChoice, ChatCompletionMessageContent, ChatCompletionResponseMessage, FinishReason,
         };
         NvCreateChatCompletionResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionResponse {
                 choices: vec![ChatChoice {
                     index: 0,
                     #[allow(deprecated)]
                     message: ChatCompletionResponseMessage {
                         content: Some(ChatCompletionMessageContent::Text(text.into())),
-                        role: dynamo_async_openai::types::Role::Assistant,
+                        role: dynamo_protocols::types::Role::Assistant,
                         tool_calls: None,
                         refusal: None,
                         reasoning_content: None,
@@ -1622,7 +1618,7 @@ thinking
 
     #[test]
     fn test_include_logprobs_kept_when_requested() {
-        use dynamo_async_openai::types::responses::IncludeEnum;
+        use dynamo_protocols::types::responses::IncludeEnum;
 
         let chat_resp = make_chat_resp_with_text("hello");
         let params = ResponseParams {
@@ -1650,7 +1646,7 @@ thinking
 
     #[test]
     fn test_truncation_auto_echoed_back() {
-        use dynamo_async_openai::types::responses::Truncation;
+        use dynamo_protocols::types::responses::Truncation;
 
         let chat_resp = make_chat_resp_with_text("hello");
         let params = ResponseParams {

lib/llm/src/protocols/openai/responses/stream_converter.rs

@@ -12,7 +12,7 @@
 use std::time::{SystemTime, UNIX_EPOCH};
 
 use axum::response::sse::Event;
-use dynamo_async_openai::types::responses::{
+use dynamo_protocols::types::responses::{
     AssistantRole, FunctionToolCall, InputTokenDetails, Instructions, OutputContent, OutputItem,
     OutputMessage, OutputMessageContent, OutputStatus, OutputTextContent, OutputTokenDetails,
     Response, ResponseCompletedEvent, ResponseContentPartAddedEvent, ResponseContentPartDoneEvent,
@@ -24,7 +24,7 @@ use dynamo_async_openai::types::responses::{
 };
 use uuid::Uuid;
 
-use dynamo_async_openai::types::ChatCompletionMessageContent;
+use dynamo_protocols::types::ChatCompletionMessageContent;
 
 use super::ResponseParams;
 use crate::protocols::openai::chat_completions::NvCreateChatCompletionStreamResponse;
@@ -673,7 +673,7 @@ fn get_event_type(event: &ResponseStreamEvent) -> &'static str {
 mod tests {
     use super::*;
     use crate::protocols::unified::ResponsesContext;
-    use dynamo_async_openai::types::{
+    use dynamo_protocols::types::{
         ChatChoiceStream, ChatCompletionMessageContent, ChatCompletionMessageToolCallChunk,
         ChatCompletionStreamResponseDelta, ChatCompletionToolType, FunctionCallStream,
     };
@@ -704,7 +704,7 @@ mod tests {
     ) -> NvCreateChatCompletionStreamResponse {
         #[allow(deprecated)]
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: "chat-1".into(),
                 choices: vec![ChatChoiceStream {
                     index: 0,
@@ -742,7 +742,7 @@ mod tests {
     fn text_chunk(text: &str) -> NvCreateChatCompletionStreamResponse {
         #[allow(deprecated)]
         NvCreateChatCompletionStreamResponse {
-            inner: dynamo_async_openai::types::CreateChatCompletionStreamResponse {
+            inner: dynamo_protocols::types::CreateChatCompletionStreamResponse {
                 id: "chat-1".into(),
                 choices: vec![ChatChoiceStream {
                     index: 0,