Dynamo v1.4.1

Dynamo v1.4.1 - Release Notes

Summary

Dynamo v1.4.1 is a patch release on top of v1.4.0. It adds OpenAI-compatible classify and pooling endpoints to the Dynamo Frontend with vLLM worker support, and fixes Dynamo Router overload recovery, logprob_token_ids passthrough, and a vLLM-Omni NIXL object transfer failure.

Base Branch: release/1.4.1

Features & Improvements

• Classify and Pooling Endpoints: Added OpenAI-compatible /v1/classify and /v1/pooling endpoints spanning the request and response protocols, model discovery, HTTP serving with metrics, and vLLM pooling-family worker activation (#13560). Pooling responses support the vLLM-compatible float, base64, bytes, and bytes_only formats including binary metadata, and truncate_prompt_tokens applies to pre-tokenized inputs using vLLM truncation semantics.

Bug Fixes

• Router Overload Mark Recovery: Fixed a Dynamo Router condition where a worker marked overloaded on the request path stayed marked indefinitely (#13432). ResourceExhausted marks a worker overloaded in the routing client, outside the monitor's cached metric set; when that cached set was empty, the next healthy load observation also computed an empty set and the monitor suppressed publication, stranding the mark. Request-path backpressure now flags the state as changed so the next load observation republishes the authoritative set.
• Logprob Token IDs Passthrough: Fixed the Dynamo Frontend rejecting logprob_token_ids with 400 Unsupported parameter(s) before the request reached the worker (#13561). The field pins the vocabulary ids whose logprobs appear in the response independent of top-k rank, which is the primitive classifier-style scoring depends on when a label would not otherwise surface in top-k. Callers should note that vLLM requires top_logprobs to match the number of pinned ids.
• NIXL Object Transfer Buffer: Fixed serialized NIXL object transfers failing in vLLM-Omni when PyTorch warned that torch.frombuffer received immutable bytes (#13610). Serialized bytes are copied into a writable buffer before the call and the now-redundant tensor clone is removed; the raw tensor transfer path is unchanged.

Version Upgrades

• Go Module Alignment: Upgraded golang.org/x/net to v0.56.0, golang.org/x/text to v0.39.0 and google.golang.org/grpc to v1.82.1 across the Dynamo Operator and inference-gateway EPP modules, and moved both to the Go v1.26.6 toolchain (#13614).

Documentation

• Classify and Pooling API Reference: Regenerated the Python API reference so the published documentation reflects the Classify and Pooling endpoint types and their supports_* accessors (#13617).
• Structural Tag Documentation Links: Fixed two structural-tag documentation links that returned 404 after xgrammar moved its documentation site to a versioned path (#13314).

Key Dependencies

Backend runtime versions are unchanged from v1.4.0:

Dynamo	SGLang	TensorRT-LLM	vLLM	NIXL	UCX
v1.4.1	v0.5.16	v1.3.0rc22	v0.26.0	v1.3.2 (vLLM) / v1.3.1 (TensorRT-LLM) / v1.3.0 (SGLang)	v1.21.0

CUDA Variants

Backend	CUDA 12	CUDA 13
SGLang	N/A	13.0
TensorRT-LLM	N/A	13.1
vLLM	N/A	13.0

Dynamo Ecosystem

AIConfigurator	AIPerf	Grove	ModelExpress
v0.11.0	v0.10.0	v0.1.0-alpha.12-rc1	v0.5.0

The following Go module dependencies changed in this release:

Module	Package	From	To
deploy/operator	google.golang.org/grpc	v1.79.3	v1.82.1
deploy/operator	Go toolchain	v1.26.5	v1.26.6
deploy/inference-gateway/epp	golang.org/x/net	v0.52.0	v0.56.0
deploy/inference-gateway/epp	golang.org/x/text	v0.35.0	v0.39.0
deploy/inference-gateway/epp	google.golang.org/grpc	v1.80.0	v1.82.1
deploy/inference-gateway/epp	Go toolchain	v1.26.3	v1.26.6

Known Issues

• No new issues filed against this release. Targeted regression on the v1.4.1 release candidate completed with no defects filed and no open blockers.
• Carried forward from v1.4.0: The known issues documented in the v1.4.0 release notes still apply.

---