Name the dev dependency groups after the vLLM line they install

.github/workflows/ci.yaml

@@ -60,7 +60,12 @@ jobs:
           enable-cache: true
           python-version: ${{ matrix.python-version }}
 
-      - run: uv sync --frozen --group dev --extra hf --extra compose
+      # `dev-vllm26`, not `dev`: the dev groups are now named after the vLLM line
+      # they install, because the GPU harness keys a stale 0.19.x version assertion
+      # off the bare name `dev` (see the comment in pyproject.toml). Nothing here is
+      # vLLM-version-specific -- it only needs pytest plus the CPU deps -- but there
+      # is one dev group per line and this is the one the lockfile resolves.
+      - run: uv sync --frozen --group dev-vllm26 --extra hf --extra compose
 
       - name: Run CPU tests
         run: |
@@ -77,8 +82,16 @@ jobs:
             -v -s --tb=short -x \
             --cov=granite_switch --cov-report=xml
 
+      # COVERAGE REPORTING MUST NOT GATE A PR. `fail_ci_if_error: false` already
+      # declared that intent, but v4 can fail the step *before* it consults that
+      # flag -- it exits nonzero on a token/rate-limit problem of its own, which is
+      # what turns a fully green CPU suite red (the "Run CPU tests" step passes on
+      # both 3.11 and 3.12; only this step fails). `continue-on-error` enforces the
+      # intent at the workflow level, where the action cannot override it. v5 is the
+      # supported line; v4 is deprecated.
       - name: Upload coverage
-        uses: codecov/codecov-action@v4
+        uses: codecov/codecov-action@v5
+        continue-on-error: true
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
           files: coverage.xml

.github/workflows/gpu-tests.yaml

@@ -110,12 +110,21 @@ jobs:
     # conflicting groups), so each gets its own cluster pod. Only the dev* groups
     # used because the bare vllm26/vllm27 groups omit pytest. The label encodes
     # the vLLM minor the leg installs (the GPU harness asserts it matches).
+    #
+    # `dep_group` MUST NOT be the bare `dev`. The harness looks this string up in a
+    # table of its own pairing each name with a job tag and a vLLM line it asserts;
+    # `dev`'s entry is frozen at 0.19.x ("dep group: dev (job tag v19, expect vllm
+    # 0.19.x)"). With `dev` on the 0.26 line since #130, this leg installed 0.26
+    # correctly and then died on that entry before pytest ran -- "FATAL expected
+    # vllm 0.19.x for group dev but got 0.26.0" -- red on every run since. A name
+    # the table does not know is not an error; it falls back to "expect vllm any.x"
+    # (which is why the dev-vllm27 leg was never affected). See pyproject.toml.
     strategy:
       fail-fast: false   # a vllm26 failure must not hide the vllm27 result
       matrix:
         include:
           - label: vllm26
-            dep_group: dev
+            dep_group: dev-vllm26
           - label: vllm27
             dep_group: dev-vllm27
     runs-on: [self-hosted, gpu]

CONTRIBUTING.md

@@ -19,10 +19,12 @@ Or via pip: `pip install uv`
    ```bash
    git clone https://github.com/<your-username>/granite-switch.git
    cd granite-switch
-   uv sync --group dev
+   uv sync --group dev-vllm26
    ```
-   > **On macOS (or any machine without a CUDA GPU):** the `dev` group pulls in vLLM + CUDA,
-   > which have no macOS wheels, so `uv sync --group dev` fails. Install the CPU-only subset
+   > The dev groups are named after the vLLM line they install (`dev-vllm26`,
+   > `dev-vllm27`); there is no bare `dev` group. Pick the line you want to work against.
+   > **On macOS (or any machine without a CUDA GPU):** these groups pull in vLLM + CUDA,
+   > which have no macOS wheels, so `uv sync --group dev-vllm26` fails. Install the CPU-only subset
    > instead — enough for unit, HF, and compose work:
    > ```bash
    > uv sync --frozen --no-default-groups --extra hf --extra compose

docs/AUDIO.md

@@ -30,7 +30,7 @@ with an `--asr-model` your installed transformers supports (for example
 uv sync --extra vllm --extra audio     # or --extra vllm27 --extra audio
 
 # Development / running the test suite (the dev groups include audio already)
-uv sync --group dev                    # vLLM 0.26.x
+uv sync --group dev-vllm26             # vLLM 0.26.x
 uv sync --group dev-vllm27             # vLLM 0.27.x
 ```
 

pyproject.toml

@@ -58,26 +58,60 @@ markers = [
 [dependency-groups]
 vllm26     = ["vllm>=0.26.0,<0.27.0"]
 vllm27     = ["vllm>=0.27.0,<0.28.0"]
+# NO DEV GROUP MAY BE NAMED A BARE `dev`. The GPU cluster harness looks the
+# `--dep-group` name up in a table of its own that pairs each name with a job tag
+# and the vLLM line it then ASSERTS the pod installed. Observed from run logs:
+#
+#     dep group: dev          (job tag v19,         expect vllm 0.19.x)
+#     dep group: dev-vllm20   (job tag v20,         expect vllm 0.20.x)
+#     dep group: dev-vllm27   (job tag dev-vllm27,  expect vllm any.x)
+#
+# Note the third line: an UNKNOWN name is not an error -- it falls back to
+# `tag = the name itself, expect any.x`, i.e. no assertion at all. Only names the
+# table knows carry a version claim, and `dev`'s claim was frozen at 0.19.x.
+#
+# So once #130 moved `dev` onto the 0.26 line, the leg installed 0.26 correctly and
+# then died on that stale entry, before pytest ever started:
+#
+#     torch 2.11.0+cu130 vllm 0.26.0
+#     FATAL expected vllm 0.19.x for group dev but got 0.26.0
+#
+# while `dev-vllm27`, absent from the table, sailed through on `any.x`. Renaming to
+# `dev-vllm26` therefore fixes the leg unconditionally -- it stops addressing the
+# one stale entry and lands on the no-assertion fallback. It does NOT depend on the
+# harness parsing "26" out of the name; it cannot, as the table is explicit.
+#
+# TRADE-OFF, AND THE REAL FOLLOW-UP: `any.x` means no leg asserts its vLLM line any
+# more, so a venv that silently resolved the wrong line would no longer be caught
+# here. The fix is for the harness table to gain real `dev-vllm26/27/30` entries;
+# until it does, the per-line `vllm26`/`vllm27` bounds below are what pin the line.
+#
+# This is a rename, not a new group: `dev` is gone rather than aliased, so the name
+# carrying the stale 0.19.x claim cannot be passed by accident. Adding a parallel
+# `dev-vllm26` and keeping `dev` was tried and rejected -- an extra conflicting
+# group multiplies uv's resolution-marker space (+3093 net lines of uv.lock for an
+# identical 275-package resolution), whereas the rename is diff-balanced.
+#
 # `audio` is included so the audio tests can actually run: the ASR path needs
 # vLLM's audio deps (av/soundfile/resampy) at runtime, and no group pulled them
 # in before (integration tests failed with ModuleNotFoundError on a synced pod).
-dev        = ["pytest", "pytest-cov", { include-group = "vllm26" }, "granite-switch[hf,compose,audio]"]
+dev-vllm26 = ["pytest", "pytest-cov", { include-group = "vllm26" }, "granite-switch[hf,compose,audio]"]
 dev-vllm27 = ["pytest", "pytest-cov", { include-group = "vllm27" }, "granite-switch[hf,compose,audio]"]
-test       = ["pytest", "pytest-cov", "bitsandbytes", "optimum-quanto", { include-group = "dev" }]
+test       = ["pytest", "pytest-cov", "bitsandbytes", "optimum-quanto", { include-group = "dev-vllm26" }]
 
 [tool.uv]
 default-groups = ["vllm26"]
 conflicts = [
     # group-vs-group
     [{ group = "vllm26" },     { group = "vllm27" }],
-    [{ group = "dev" },        { group = "vllm27" }],
-    [{ group = "dev" },        { group = "dev-vllm27" }],
+    [{ group = "dev-vllm26" }, { group = "vllm27" }],
+    [{ group = "dev-vllm26" }, { group = "dev-vllm27" }],
     [{ group = "dev-vllm27" }, { group = "vllm26" }],
     # group-vs-extra
     [{ group = "vllm26" },     { extra = "vllm27" }],
     [{ group = "vllm27" },     { extra = "vllm" }],
     [{ group = "vllm27" },     { extra = "tutorials" }],
-    [{ group = "dev" },        { extra = "vllm27" }],
+    [{ group = "dev-vllm26" }, { extra = "vllm27" }],
     [{ group = "dev-vllm27" }, { extra = "vllm" }],
     [{ group = "dev-vllm27" }, { extra = "tutorials" }],
     # extra-vs-extra