From c458d21d077d28bff36565204732405a93d3661d Mon Sep 17 00:00:00 2001 From: Tzu-Yu Lee Date: Sun, 11 May 2025 03:27:59 +0800 Subject: [PATCH] feat(trtllm): add new finish reasons Add new finish reasons introduced in TensorRT-LLM v0.16.0. --- backends/trtllm/csrc/ffi.hpp | 4 ++++ backends/trtllm/src/lib.rs | 8 ++++++++ 2 files changed, 12 insertions(+) diff --git a/backends/trtllm/csrc/ffi.hpp b/backends/trtllm/csrc/ffi.hpp index 840614bb..a877df5a 100644 --- a/backends/trtllm/csrc/ffi.hpp +++ b/backends/trtllm/csrc/ffi.hpp @@ -42,6 +42,10 @@ namespace huggingface::tgi::backends::trtllm { return finish_reason_t::kEND_ID; case tle::FinishReason::kLENGTH: return finish_reason_t::kLENGTH; + case tle::FinishReason::kTIMED_OUT: + return finish_reason_t::kTIMED_OUT; + case tle::FinishReason::kCANCELLED: + return finish_reason_t::kCANCELLED; default: std::unreachable(); } diff --git a/backends/trtllm/src/lib.rs b/backends/trtllm/src/lib.rs index 08507256..52e48f91 100644 --- a/backends/trtllm/src/lib.rs +++ b/backends/trtllm/src/lib.rs @@ -24,6 +24,14 @@ mod ffi { /// The request finished because the maximum number of tokens was reached. #[cxx_name = "kLENGTH"] MaxLength = 3u8, + + #[cxx_name = "kTIMED_OUT"] + /// The request finished because it got timed out (via the mAllotedTime parameter) + TimedOut = 4u8, + + #[cxx_name = "kCANCELLED"] + /// The request was cancelled by calling cancelRequest. + Cancelled = 5u8, } /// Struct used as shared type between rust and C++ to represent the result