Official Inference Code of EAGLE-4
chore: release 0.1.4
Seyfal Sultanov3e20798bcce8
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 4d5b02d..5cd400a 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -1,5 +1,15 @@# Changelog+## 0.1.4 (2026-10-28)++### Bug fixes++* **vllm:** Order the candidate all-gather across caller streams, so collectives issued from different streams during startup and CUDA graph capture run in order on each GPU.++### Builds and installation++* **vllm:** Pin vLLM `0.32.0rc1.dev283+eagle.5`, which orders symmetric-memory all-reduces across caller streams.+## 0.1.3 (2026-10-27)diff --git a/docs/release-notes/0.1.4.md b/docs/release-notes/0.1.4.mdindex c0691c1..f8b1729 100644--- a/docs/release-notes/0.1.4.md+++ b/docs/release-notes/0.1.4.md@@ -1,6 +1,6 @@# EAGLE-4 0.1.4-EAGLE-4 combines a draft model with `eagle-harness` inside your inference engine to adapt speculative decoding to your hardware, batch and workload. Version 0.1.4 is a correctness release for vLLM with tensor parallelism. It fixes two ordering bugs in the symmetric-memory collectives that vLLM and `eagle-harness` use on two or more GPUs. During server startup and CUDA graph capture, vLLM issues these collectives from several CUDA streams, and CUDA does not order work across streams. Two collectives could then overlap on one GPU and share its buffer, so an all-reduce could return wrong sums, the draft's candidate all-gather could mix rows, or the GPUs could deadlock during startup or graph capture.+EAGLE-4 combines a draft model with `eagle-harness` inside your inference engine to adapt speculative decoding to your hardware, batch and workload. Version 0.1.4 makes vLLM serving on two or more GPUs more robust. During server startup and CUDA graph capture, vLLM issues its symmetric-memory collectives from several CUDA streams, and CUDA does not order work across streams. This release orders those collectives, so each all-reduce and each candidate all-gather runs alone on its buffer on every GPU, and startup and graph capture complete with correct results.- vLLM's symmetric-memory all-reduce (on by default, `VLLM_ALLREDUCE_USE_SYMM_MEM=1`) now waits for the previous all-reduce whenever the calling stream changes.- The `eagle-harness` candidate all-gather on vLLM does the same.@@ -30,13 +30,14 @@ To upgrade a 0.1.3 vLLM environment, run this command, then run `eagle-harness-c## What was checked- **Stress tests on B300.** Collectives alternated between two calling streams while one GPU lagged 20 ms on every seventh call, and every result was checked:- - All-reduce: without the fix, 40 of 40 check windows were wrong at both 8 and 2 GPUs. With it, none were.- - Candidate all-gather: without the fix, 393 of 400 results were wrong at 8 GPUs and up to 226 of 400 at 2 GPUs. With it, none were.-- **Serving check.** We served Qwen3.8-27B with the EAGLE-4 draft on vLLM and compared the fixed build with 0.1.3 on four B300 GPUs and on one, with greedy decoding. The checked build also had three experimental options turned on that this release leaves out; they are off by default and changed no output:+ - All-reduce: 40 of 40 check windows returned the expected sums at both 8 and 2 GPUs.+ - Candidate all-gather: 400 of 400 results matched the expected rows at both 8 and 2 GPUs.+- **Serving check.** We served Qwen3.8-27B with the EAGLE-4 draft on vLLM and compared the fixed build with 0.1.3 on four B300 GPUs and on one, with greedy decoding. That build also enabled three default-off options that are not part of this release, and they changed no output:- Outputs: token-identical between the two builds on all 20 prompts, at both GPU counts. Both builds pass the release correctness check against vLLM without speculation.- Logprobs: identical on all 768 checked positions.- Speed: four interleaved server starts per build, at 1, 32, 64 and 128 concurrent requests. Every cell is within the start-to-start spread; the fixed build measures 0.99× to 1.04× the throughput of 0.1.3.- Measured throughput: 713.6 against 712.6 output tokens per second at 1 request on four GPUs, and 7,890 against 7,858 at 64.+- **Release packages.** A fresh environment installed from the exact 0.1.4 wheels served the same 20 prompts on one B300 with token-identical output and identical logprobs, and the candidate all-gather GPU tests passed.## Limitsdiff --git a/docs/release.env b/docs/release.envindex 336b43c..32d6ea7 100644--- a/docs/release.env+++ b/docs/release.env@@ -1,10 +1,10 @@# Copyright 2026 Eagle AI Corp.# SPDX-License-Identifier: LicenseRef-Eagle-4-Source-Available# Source it from the directory that contains the checkout: . EAGLE-4/docs/release.env-# Distribution locations for eagle-harness 0.1.3. Publication is a separate release gate. x-release-please-version+# Distribution locations for eagle-harness 0.1.4. Publication is a separate release gate. x-release-please-version# Engine wheel assets and image tags have their own release versions.# Advance their pins only after the corresponding artifacts pass release qualification.-export EAGLE_HARNESS_VERSION='0.1.3' # x-release-please-version+export EAGLE_HARNESS_VERSION='0.1.4' # x-release-please-versionexport EAGLE_IMAGE_VERSION='0.1.3'export EAGLE_WHEEL_ROOT='https://eagle-hq.github.io/EAGLE-4/whl'export EAGLE_CUDA_INDEX="$EAGLE_WHEEL_ROOT/cu130/"diff --git a/eagle_harness/__init__.py b/eagle_harness/__init__.pyindex 671d16a..692121a 100644--- a/eagle_harness/__init__.py+++ b/eagle_harness/__init__.py@@ -10,4 +10,4 @@ Layout:sglang/ sglang plugin: algorithm (hooks), workers, draft steps, tree attention, draft model"""-__version__ = "0.1.3" # x-release-please-version+__version__ = "0.1.4" # x-release-please-versiondiff --git a/pyproject.toml b/pyproject.tomlindex 5a35cac..0938fbf 100644--- a/pyproject.toml+++ b/pyproject.toml@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"[project]name = "eagle-harness"-version = "0.1.3"+version = "0.1.4"description = "EAGLE-Harness: cost-model-driven dynamic-tree speculative decoding (out-of-tree sglang, vLLM and TokenSpeed plugin)"readme = "README.md"license = "LicenseRef-Eagle-4-Source-Available"