release: 0.3.2 (version bump, CHANGELOG, image tag 0.3.2)

CHANGELOG.md

@@ -10,6 +10,31 @@ Versioning follows [Semantic Versioning 2.0.0](https://semver.org/).
 
 ## [Unreleased]
 
+## [0.3.2] — 2026-10-29 — Per-call acceptance rate
+
+Development status stays Alpha. A bug-fix release with no new measurement:
+the decoding path and its speed are unchanged, so the README tables remain
+the 0.3.0 release run (bundle `3c43e08`), with the acceptance column now
+computed from that run's stored step counts.
+
+### Highlights
+- `GenerationResult.acceptance_rate` reports the call that returned it.
+  Before, every call after the first on an engine reported a running mean
+  over all earlier calls, warm-ups included.
+- The README acceptance column and Figure 9 of `docs/lessons.md` are computed
+  from the stored per-step counts.
+
+### Upgrade notes
+- Nothing to change in your code. If you logged `acceptance_rate` from
+  repeated calls on one engine with 0.3.1 or earlier, those values were
+  running means; per-call values are accepted / proposed summed over that
+  call's `arm_steps`, which were always correct.
+
+### Known issues
+- Unchanged from 0.3.1 (first call compiles; fp16 greedy exact ties; batch
+  > 1 and other GPUs or model pairs not measured; bandit experimental; the
+  CUDA-graph "distinct sizes" warning).
+
 ### Fixed
 - `GenerationResult.acceptance_rate` now covers only the call that returned
   it, as documented. The engine created its acceptance tracker once and never

CITATION.cff

@@ -35,8 +35,8 @@ abstract: >-
   UCB1/Thompson draft selector is experimental: measured once over seven
   arms, it did not beat the best fixed arm. The Triton and assembly
   kernels are experimental and not on the generation critical path.
-version: 0.3.1
-date-released: "2026-10-26"
+version: 0.3.2
+date-released: "2026-10-29"
 license: Apache-2.0
 repository-code: "https://github.com/Mattral/FlashSpec"
 url: "https://flashspec.readthedocs.io"

README.md

@@ -45,8 +45,8 @@ You need Python 3.11 or later, PyTorch 2.2 or later and transformers 4.40 or lat
 **Container image (GPU-oriented).** Each release is also published as a server image on GitHub Container Registry. It is built from the PyPI release on CUDA 13.0, with the torch build the GPU results were measured with, and it runs as a non-root user. It needs an NVIDIA GPU and the NVIDIA Container Toolkit; it starts on CPU too, but CPU is not where FlashSpec is faster. The default command serves the measured pair in fp16 on port 8000.
 
 ```bash
-docker pull ghcr.io/mattral/flashspec:0.3.1
-docker run --gpus all -p 8000:8000 ghcr.io/mattral/flashspec:0.3.1
+docker pull ghcr.io/mattral/flashspec:0.3.2
+docker run --gpus all -p 8000:8000 ghcr.io/mattral/flashspec:0.3.2
 ```
 
 ## Quickstart
@@ -153,7 +153,7 @@ python -m flashspec.serve --target Qwen/Qwen2.5-1.5B-Instruct --drafter Qwen/Qwe
 Or with the container image. Arguments replace its default command, so pass `--host 0.0.0.0`. The volume keeps downloaded checkpoints between runs:
 
 ```bash
-docker run --gpus all -p 8000:8000 -v flashspec-hf:/home/flashspec/.cache/huggingface ghcr.io/mattral/flashspec:0.3.1 \
+docker run --gpus all -p 8000:8000 -v flashspec-hf:/home/flashspec/.cache/huggingface ghcr.io/mattral/flashspec:0.3.2 \
     --target Qwen/Qwen2.5-1.5B-Instruct --drafter Qwen/Qwen2.5-0.5B-Instruct \
     --device cuda:0 --dtype float16 --temperature 0 --host 0.0.0.0
 ```
@@ -328,7 +328,7 @@ With `compile` on, both models run on static KV caches and the fixed-shape decod
 
 ## Status and limitations
 
-FlashSpec is alpha, version 0.3.1.
+FlashSpec is alpha, version 0.3.2.
 
 - **Measured scope.** One model pair (`Qwen/Qwen2.5-1.5B-Instruct` verified by `Qwen/Qwen2.5-0.5B-Instruct`), one Tesla T4, fp16, batch 1, 10 MT-Bench and 10 HumanEval prompts, run from `main` at git `3c43e08` with the defaults `compile="auto"` and `drafter_strategy="auto"` (which ran `composite-g6-f2`). No other GPU, model pair or batch size has been measured. Every H100 number in the docs is a design target.
 - **The first call compiles.** On CUDA the first `generate` compiles the forwards and captures CUDA graphs, which can take minutes; later calls reuse them. The tables exclude that call.

deploy/Dockerfile

@@ -1,6 +1,6 @@
 # FlashSpec server image: the released PyPI package on a CUDA 13.0 base.
 #
-#   docker build -f deploy/Dockerfile --build-arg FLASHSPEC_VERSION=0.3.1 -t flashspec deploy/
+#   docker build -f deploy/Dockerfile --build-arg FLASHSPEC_VERSION=0.3.2 -t flashspec deploy/
 #   docker run --gpus all -p 8000:8000 flashspec
 #
 # GPU-oriented: torch is the CUDA 13.0 build FlashSpec's GPU results were

deploy/docker-compose.yml

@@ -7,7 +7,7 @@
 
 services:
   flashspec:
-    image: ghcr.io/mattral/flashspec:0.3.1
+    image: ghcr.io/mattral/flashspec:0.3.2
     ports:
       - "8000:8000"
     volumes:

deploy/k8s/flashspec-deployment.yaml

@@ -4,7 +4,7 @@ metadata:
   name: flashspec
   labels:
     app: flashspec
-    version: "0.3.1"
+    version: "0.3.2"
 spec:
   replicas: 1
   selector:
@@ -19,7 +19,7 @@ spec:
         - name: flashspec
           # GPU-oriented image; the default command serves Qwen2.5-1.5B-Instruct
           # verified by Qwen2.5-0.5B-Instruct in fp16 on port 8000.
-          image: ghcr.io/mattral/flashspec:0.3.1
+          image: ghcr.io/mattral/flashspec:0.3.2
           imagePullPolicy: IfNotPresent
           ports:
             - containerPort: 8000

docs/install.md

@@ -26,8 +26,8 @@ git clone https://github.com/Mattral/FlashSpec && cd FlashSpec
 pip install -e ".[dev,gpu]"   # omit ",gpu" on Windows/macOS
 
 # Container image (Linux + NVIDIA GPU), published for each release:
-docker pull ghcr.io/mattral/flashspec:0.3.1
-docker run --gpus all -p 8000:8000 ghcr.io/mattral/flashspec:0.3.1
+docker pull ghcr.io/mattral/flashspec:0.3.2
+docker run --gpus all -p 8000:8000 ghcr.io/mattral/flashspec:0.3.2
 ```
 
 **The container image** (`ghcr.io/mattral/flashspec:<version>` and `:latest`) is
@@ -48,7 +48,7 @@ built by `.github/workflows/container.yml` when a release is published:
 To build it yourself from a released version:
 
 ```bash
-docker build -f deploy/Dockerfile --build-arg FLASHSPEC_VERSION=0.3.1 -t flashspec deploy/
+docker build -f deploy/Dockerfile --build-arg FLASHSPEC_VERSION=0.3.2 -t flashspec deploy/
 ```
 
 A development checkout, by platform:

flashspec/__init__.py

@@ -42,5 +42,5 @@
     "register",
 ]
 
-__version__ = "0.3.1"
+__version__ = "0.3.2"
 __author__ = "Min Htet Myet (Mattral)"

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "flashspec"
-version = "0.3.1"
+version = "0.3.2"
 description = "Lossless speculative decoding for HuggingFace causal LMs, with a static-cache CUDA-graph fast path; bandit draft selection and Triton kernels are experimental"
 readme = {file = "README.md", content-type = "text/markdown"}
 license = { file = "LICENSE" }