sie-server 0.7.0__tar.gz → 0.7.2__tar.gz

{sie_server-0.7.0 → sie_server-0.7.2}/.gitignore

@@ -142,10 +142,8 @@ celerybeat.pid
 
 # Environments
 .env
-# Managed-service cloud provisioning secrets (cloud-provision / deploy/cloud).
-# Vendor credentials live in the env or a gitignored secrets file, never the
-# repo (audit §1/§14). Belt-and-braces so a local copy can never be committed.
-*.sie-cloud-secrets.env
+# Local service credential bundles must never be committed.
+*-secrets.env
 .env.cloud
 .venv
 env/
@@ -244,33 +242,6 @@ override.tf.json
 # tfvars files may contain secrets
 *.tfvars
 *.tfvars.json
-# Reviewed worker-AMI release contracts contain only immutable public package,
-# AMI, and Terraform backend coordinates; the release workflow accepts only
-# committed files from this exact internal path.
-!deploy/terraform/aws/internal-examples/cloud-benchmark-runner/worker-ami/releases/*.tfvars.json
-# Exception: this edge tfvars carries only a Secrets Manager ARN (a resource
-# identifier, not a secret) so the staging-us edge apply needs no manual -var.
-!deploy/cloud/terraform/aws-edge/staging-us.tfvars
-# This dev adoption contract contains only public AWS resource identifiers; it
-# is checked in so a reviewed plan cannot silently target a different orphan
-# VPC/listener through an operator-local tfvars file.
-!deploy/cloud/terraform/dev-controlplane-https-adoption/dev-us.tfvars
-!deploy/cloud/terraform/dev-controlplane-https-adoption/.terraform.lock.hcl
-# Managed release roots commit their exact provider selections. Release
-# preflights and applies use these read-only instead of selecting a newer
-# compatible provider during a long-running deployment.
-!deploy/cloud/terraform/release-infrastructure/.terraform.lock.hcl
-!deploy/cloud/terraform/aws-controlplane/.terraform.lock.hcl
-!deploy/cloud/terraform/aws-edge/.terraform.lock.hcl
-!deploy/cloud/terraform/s3-state/.terraform.lock.hcl
-!deploy/cloud/terraform/vercel/.terraform.lock.hcl
-!deploy/cloud/terraform/console-preview/.terraform.lock.hcl
-!deploy/cloud/terraform/examples/dev/.terraform.lock.hcl
-!deploy/cloud/terraform/examples/dev-eu/.terraform.lock.hcl
-!deploy/cloud/terraform/examples/staging/.terraform.lock.hcl
-!deploy/cloud/terraform/examples/staging-eu/.terraform.lock.hcl
-!deploy/cloud/terraform/examples/prod-us/.terraform.lock.hcl
-# Keep .terraform.lock.hcl for reproducibility (provider versions)
 
 # Node.js
 node_modules/
@@ -294,12 +265,7 @@ tmp/
 .tmp/
 .local/
 
-.requirements-modal.txt
-.mpa-i6pn-*-requirements-modal.txt
 b6_results.json
-# Raw dev-cloud load-test reports are local artifacts, not source-controlled baselines.
-packages/sie_cloud/bench/results/dev-us-*/
-
 # Personal mise overrides (per-developer toolchain tweaks, e.g. rustup
 # component naming differences across rustup versions).
 mise.local.toml

{sie_server-0.7.0 → sie_server-0.7.2}/CONTRIBUTING.md

@@ -14,12 +14,12 @@ mise trust && mise install
 ## Running Tests
 
 ```bash
-mise run test packages/sie_server
+mise run test -- packages/sie_server
 ```
 
 ## Code Style
 
-- Format with `mise run lint -f`
+- Format with `mise run lint -- -f`
 - Type check with `mise run typecheck`
 
 ## Pull Requests

{sie_server-0.7.0 → sie_server-0.7.2}/Dockerfile.cuda12

@@ -165,7 +165,7 @@ RUN .venv/bin/python -m sie_server.cli resolve-deps \
 #
 # Per-bundle CUDA wheel index. SGLang's sgl-kernel/flashinfer wheels live on the
 # docs.sglang.ai cu129 index (not PyPI), so the SGLang-family bundles add it.
-# The gemma bundle (sglang 0.5.13 / torch 2.11 / cu130, CUDA 13) is built by the
+# The sglang-cu130 bundle (sglang 0.5.13 / torch 2.11 / cu130, CUDA 13) is built by the
 # separate ``cuda13`` platform — see Dockerfile.cuda13 — not here.
 RUN --mount=type=cache,target=/root/.cache/pip \
     --mount=type=cache,target=/root/.cache/uv \
@@ -216,9 +216,9 @@ ENV CUDA_HOME=/usr/local/cuda \
     LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
 FROM runtime-sglang AS runtime-sglang-embedding
 FROM runtime-sglang AS runtime-sglang-vision-extract
-# NOTE: the gemma bundle (CUDA 13 / sglang 0.5.13 / torch 2.11 cu130) is built by
+# NOTE: the sglang-cu130 bundle (CUDA 13 / sglang 0.5.13 / torch 2.11 cu130) is built by
 # the separate ``cuda13`` platform — packages/sie_server/Dockerfile.cuda13 — so
-# this CUDA-12 image never produces a (mislabelled) CUDA-13 ``cuda12-gemma``
+# this CUDA-12 image never produces a (mislabelled) CUDA-13 ``cuda12-sglang-cu130``
 # image. See #1415.
 
 FROM runtime-${BUNDLE} AS runtime
@@ -294,7 +294,8 @@ COPY --link --from=base --chown=1000:1000 /app/bundles /app/bundles
 # A `.pth` file in the shared venv puts /app/bundle-libs at sys.path[0] at startup.
 COPY --link --from=builder --chown=1000:1000 /app/bundle-libs /app/bundle-libs
 
-RUN mkdir -p /app/evals /app/.cache/huggingface && chown -R sie:sie /app/evals /app/.cache
+RUN mkdir -p /app/evals /app/.cache/huggingface /app/.cache/sie-kernels \
+    && chown -R sie:sie /app/evals /app/.cache
 
 # Declare the BUNDLE arg only here, where it is first used (LABEL + ENV),
 # so every RUN/COPY layer above is bundle-agnostic in its cache key and the
@@ -309,6 +310,7 @@ ENV PATH="/app/.venv/bin:$PATH" \
     PYTHONUNBUFFERED=1 \
     PYTHONDONTWRITEBYTECODE=1 \
     HF_HOME=/app/.cache/huggingface \
+    SIE_SGLANG_KERNEL_CACHE_ROOT=/app/.cache/sie-kernels \
     NVIDIA_VISIBLE_DEVICES=all \
     NVIDIA_DRIVER_CAPABILITIES=compute,utility \
     SIE_BUNDLE="${BUNDLE}"

{sie_server-0.7.0 → sie_server-0.7.2}/Dockerfile.cuda13

@@ -1,13 +1,13 @@
 # syntax=docker/dockerfile:1
-# SIE Server - CUDA 13 Image (Gemma 4 / cu130 bundles)
+# SIE Server - CUDA 13 Image (SGLang cu130 bundles)
 # Build from repo root:
-#   docker build -f packages/sie_server/Dockerfile.cuda13 --build-arg BUNDLE=gemma -t sie-server:cuda13-gemma .
+#   docker build -f packages/sie_server/Dockerfile.cuda13 --build-arg BUNDLE=sglang-cu130 -t sie-server:cuda13-sglang-cu130 .
 #
-# WHY A SEPARATE PLATFORM: the gemma bundle needs sglang 0.5.13, which pins
+# WHY A SEPARATE PLATFORM: the sglang-cu130 bundle needs sglang 0.5.13, which pins
 # torch==2.11.0 (cu130) + transformers 5.8.x and JIT-compiles cu130 kernels
 # (flashinfer / tvm_ffi) through nvcc on first decode — so it requires a CUDA 13
 # devel runtime. Keeping it here (platform ``cuda13``) instead of inside
-# Dockerfile.cuda12 means a ``cuda13-gemma`` image is honestly labelled CUDA 13,
+# Dockerfile.cuda12 means a ``cuda13-sglang-cu130`` image is honestly labelled CUDA 13,
 # not a CUDA-13 image hiding under a ``cuda12-*`` tag (see #1415).
 #
 # STRUCTURE: mirrors Dockerfile.cuda12's deps→base→builder→runtime stages so the
@@ -18,18 +18,18 @@
 # cu130 bundle), so there is no per-bundle runtime indirection.
 #
 # DUAL-TORCH CARRY: the shared base venv still installs torch 2.9 (cu129) from
-# pyproject, and the gemma bundle overlays torch 2.11 (cu130) into bundle-libs
+# pyproject, and the sglang-cu130 bundle overlays torch 2.11 (cu130) into bundle-libs
 # (~+3 GB). Slimming torch out of sie-server core to drop the carry is a tracked
 # follow-up (#1415).
 
-ARG BUNDLE=gemma
+ARG BUNDLE=sglang-cu130
 ARG UV_VERSION=0.9.28
 
-# Build the internal Rust extension from checked-in source; it is not published
+# Build the native Rust extension from checked-in source; it is not published
 # to PyPI.
 FROM ghcr.io/pyo3/maturin:v1.13.3 AS audio-prep-builder
 WORKDIR /io
-# The crate is a Cargo workspace member (#2339): its pins live in the root
+# The crate is a Cargo workspace member: its pins live in the root
 # Cargo.lock, and loading the workspace needs every member's manifest+sources.
 COPY Cargo.toml Cargo.lock ./
 COPY packages/sie_audio_prep/ packages/sie_audio_prep/
@@ -119,7 +119,7 @@ RUN .venv/bin/python -c "import sie_audio_prep, torch; print(torch.__version__)"
 
 # Register bundle-libs on sys.path via a .pth executed by site.py at startup.
 # sys.path.insert(0, ...) puts bundle-libs at position 0 so bundle-specific
-# versions SHADOW the shared venv (the gemma cu130 torch/transformers/sglang
+# versions SHADOW the shared venv (the cu130 torch/transformers/sglang
 # overlay). Content is identical for every bundle, so writing it here keeps the
 # shared venv layer byte-identical.
 RUN echo "import sys; sys.path.insert(0, '/app/bundle-libs')" \
@@ -155,16 +155,16 @@ RUN .venv/bin/python -m sie_server.cli resolve-deps \
     && cat /tmp/bundle-requirements.txt
 
 # cu130 bundle deps go into a separate site-packages tree (``bundle-libs``,
-# prepended to sys.path via the .pth above). The gemma bundle needs torch 2.11
+# prepended to sys.path via the .pth above). The sglang-cu130 bundle needs torch 2.11
 # from the cu130 PyTorch index, where the only ``2.11.0`` is the ``+cu130``
 # local-version wheel; pip's multi-index resolution does not reliably prefer it
 # over PyPI's default-CUDA build, so cuda13 bundles resolve with ``uv`` and
 # ``--index-strategy unsafe-best-match`` against the cu130 PyTorch + sglang
 # indexes (so torch 2.11 + sgl-kernel/flashinfer land for cu130). This uv +
-# cu130-index resolution was validated by a manual Modal GPU session (sglang
-# 0.5.13 / torch 2.11 / cu130 in a fresh venv).
+# cu130-index resolution was validated with sglang 0.5.13, torch 2.11, and a
+# cu130 runtime in a fresh GPU environment.
 #
-# COVERAGE NOTE: the cuda13 image CI smoke (.github/workflows/gemma-image.yml)
+# COVERAGE NOTE: the cuda13 image CI smoke (.github/workflows/sglang-cu130-image.yml)
 # builds this image and asserts on CPU that the bundle-libs cu130 overlay
 # SHADOWS the base venv (import torch -> 2.11+cu130, transformers 5.x). It does
 # NOT exercise GPU init: the cu12 base venv + cu130 bundle-libs both expose the
@@ -268,7 +268,8 @@ COPY --link --from=base --chown=1000:1000 /app/bundles /app/bundles
 # Bundle-specific extras — last layer so shared layers above stay cached.
 COPY --link --from=builder --chown=1000:1000 /app/bundle-libs /app/bundle-libs
 
-RUN mkdir -p /app/evals /app/.cache/huggingface && chown -R sie:sie /app/evals /app/.cache
+RUN mkdir -p /app/evals /app/.cache/huggingface /app/.cache/sie-kernels \
+    && chown -R sie:sie /app/evals /app/.cache
 
 # Declare BUNDLE only here (LABEL + ENV) so layers above are bundle-agnostic.
 ARG BUNDLE
@@ -281,6 +282,7 @@ ENV PATH="/app/.venv/bin:$PATH" \
     PYTHONUNBUFFERED=1 \
     PYTHONDONTWRITEBYTECODE=1 \
     HF_HOME=/app/.cache/huggingface \
+    SIE_SGLANG_KERNEL_CACHE_ROOT=/app/.cache/sie-kernels \
     NVIDIA_VISIBLE_DEVICES=all \
     NVIDIA_DRIVER_CAPABILITIES=compute,utility \
     SIE_BUNDLE="${BUNDLE}"

{sie_server-0.7.0 → sie_server-0.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: sie-server
-Version: 0.7.0
+Version: 0.7.2
 Summary: Search Inference Engine - GPU inference server for search workloads
 License: Apache-2.0
 License-File: LICENSE

{sie_server-0.7.0 → sie_server-0.7.2}/README.md

@@ -67,7 +67,7 @@ auto-retries; see `packages/sie_sdk/README.md` for client-side controls.
 | Env var | Default | Effect |
 |--|--|--|
 | `SIE_MAX_BATCH_REQUESTS` | `64` | Maximum number of items per batched inference call. |
-| `SIE_MAX_BATCH_WAIT_MS` | `10` | How long the batch-former waits for additional items before dispatching. |
+| `SIE_MAX_BATCH_WAIT_MS` | `15.0` | Initial value for the adaptive first-request batch timeout. At runtime the PI batching controller steers it between `SIE_ADAPTIVE_BATCHING__MIN_WAIT_MS` and `SIE_ADAPTIVE_BATCHING__MAX_WAIT_MS`, so it is a starting point rather than a fixed wait. |
 | `SIE_MAX_CONCURRENT_REQUESTS` | `512` | Per-worker queue size; admission control returns `QUEUE_FULL` above this. |
 | `SIE_MAX_LORAS_PER_MODEL` | `10` | Maximum concurrent LoRA adapters per base model. |
 

{sie_server-0.7.0 → sie_server-0.7.2}/bundles/default.yaml

@@ -59,8 +59,17 @@ deps:
   sentence-transformers: '>=5.0,<6'
   # bge_m3_flag
   FlagEmbedding: '>=1.3'
-  # grounding_dino
-  pillow: '>=11,<12'
+  # grounding_dino. EXACT, not a range. `uv run --with-requirements` resolves the
+  # overlay live from PyPI, and `--frozen`/`--no-sync` govern only the project
+  # env, so any unbounded range here is re-resolved on every eval. The old
+  # `>=11,<12` (94430893c) was accidentally stable because the 11.x line is
+  # closed, but it contradicted sie-server's own `pillow>=12.2.0` since the
+  # security bump (27fd2063d), so every default-bundle eval served the superseded
+  # major. A bare `>=12.2.0` would fix that while making the pin float, which
+  # trades a stale-but-stable closure for a moving one.
+  # Pinned to the version the project lock holds, so overlay and lock agree.
+  # Changing this re-measures 13 gating floors (naver/v-splade-quality, docling).
+  pillow: '==12.3.0'
   # grounding_dino
   requests: ''
   # gliner, gliner_bi

sie_server-0.7.0/bundles/gemma.yaml → sie_server-0.7.2/bundles/sglang-cu130.yaml

@@ -1,15 +1,14 @@
-name: gemma
+name: sglang-cu130
 priority: 20
 default: false
 # GPU platform this bundle targets — selects packages/sie_server/Dockerfile.cuda13
-# and tags images ``cuda13-gemma`` (not ``cuda12-*``). Gemma 4 needs sglang
+# and tags images ``cuda13-sglang-cu130`` (not ``cuda12-*``). Gemma 4 needs sglang
 # 0.5.13 / torch 2.11 cu130 on a CUDA 13 devel runtime (see below). Read by the
-# docker build task (tools/mise_tasks/docker_task.py) and mirrored by the release
-# matrix (.github/release-matrix.json). Bundles without this field default to
-# ``cuda12``.
+# image build. Bundles without this field default to ``cuda12``.
 platform: cuda13
-# CUDA 13 SGLang worker image, introduced for Google Gemma 4 (see issue #1415)
-# and reused by profiles that require newer generic engine features.
+# General CUDA 13 / cu130 SGLang worker image, introduced for Google Gemma 4
+# (see issue #1415) and shared by every profile that requires this newer engine
+# and toolchain closure, including Qwen3.6 and Qwen3.8 native-window profiles.
 #
 # VALIDATED on a Modal L4 (driver 580.95.05 / CUDA 13.0) on 2026-06-18:
 # google/gemma-4-E2B-it serves end-to-end on sglang 0.5.13 — the gemma4
@@ -17,8 +16,8 @@ platform: cuda13
 # and /v1/chat/completions returns 200 (with --disable-cuda-graph, which the
 # model configs set via adapter_options.loadtime.disable_cuda_graph).
 #
-# Kept SEPARATE from the `sglang` bundle so its Qwen3.x default/non-thinking
-# profiles stay on sglang 0.5.10.post1 / torch 2.9.1. Qwen3.6-27B native-window
+# Kept SEPARATE from the `sglang` bundle so its older Qwen3.x profiles stay on
+# sglang 0.5.10.post1 / torch 2.9.1. Qwen3.6-27B native-window
 # variants opt into this image through the generic CUDA 13 adapter; their
 # thinking variants add the strict-thinking behavior. The rest of the Qwen
 # catalog remains untouched.
@@ -32,7 +31,7 @@ platform: cuda13
 # this bundle's image carries torch 2.11 (cu130) on a CUDA 13 base — built by
 # the dedicated ``cuda13`` platform (packages/sie_server/Dockerfile.cuda13),
 # selected via the ``platform: cuda13`` field above. The shared base venv still
-# ships torch 2.9.1, so the gemma image carries two torches (~+3 GB); slimming
+# ships torch 2.9.1, so the cu130 image carries two torches (~+3 GB); slimming
 # torch out of sie-server core would remove that (tracked follow-up in #1415).
 #
 # Distinct thin adapter modules scope routing here because bundle compatibility
@@ -47,7 +46,8 @@ deps:
   # sglang 0.5.13 = the validated Gemma 4 engine (0.5.11 has the activation.cuh
   # JIT bug above). Verified transitive resolution on the Modal L4 cu130 image:
   #   torch==2.11.0 (cu130), transformers==5.8.1, kernels==0.14.1,
-  #   sglang-kernel + flashinfer-python (0.5.13 builds).
+  #   sgl-deep-gemm==0.1.2, sglang-kernel==0.4.3, and
+  #   flashinfer-python==0.6.12 (the exact SGLang 0.5.13 closure).
   # These resolve from the cu130 wheel indexes at build/serve time
   # (download.pytorch.org/whl/cu130 + https://docs.sglang.ai/whl/cu130/),
   # supplied by the build (Dockerfile.cuda13 / Modal image). torch,

{sie_server-0.7.0 → sie_server-0.7.2}/bundles/sglang-embedding.yaml

@@ -20,9 +20,8 @@ deps:
   # a pre-release-only requirement. PyPI's ``flash-attn-4`` also publishes a
   # stable ``0.0.1`` placeholder, which disables uv's "allow pre-releases
   # when a package has nothing but pre-releases" rule, so a resolution that
-  # only names ``sglang`` is unsatisfiable (nightly quality-adapter 7B lanes
-  # died on this for weeks — #1684). Declaring the dep here with an explicit
-  # pre-release specifier re-enables pre-releases for exactly this package.
+  # only names ``sglang`` is unsatisfiable. Declaring the dep here with an
+  # explicit pre-release specifier re-enables pre-releases for exactly this package.
   # Runtime-inert for us: sglang's default attention backend is flashinfer;
   # fa4 only needs to be importable to satisfy the dependency.
   flash-attn-4: '==4.0.0b19'

{sie_server-0.7.0 → sie_server-0.7.2}/models/Alibaba-NLP__gte-multilingual-base.yaml

@@ -1,5 +1,6 @@
 sie_id: Alibaba-NLP/gte-multilingual-base
 hf_id: Alibaba-NLP/gte-multilingual-base
+hf_revision: 9bbca17d9273fd0d03d5725c7a4b0f6b45142062
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/BAAI__bge-reranker-large.yaml

@@ -1,5 +1,6 @@
 sie_id: BAAI/bge-reranker-large
 hf_id: BAAI/bge-reranker-large
+hf_revision: 55611d7bca2a7133960a6d3b71e083071bbfc312
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/EmergentMethods__gliner_large_news-v2.1.yaml

@@ -1,5 +1,6 @@
 sie_id: EmergentMethods/gliner_large_news-v2.1
 hf_id: EmergentMethods/gliner_large_news-v2.1
+hf_revision: 1af0cd2ce3b4a2c660849f6c3a10c7a7bac3c6c6
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/GritLM__GritLM-7B.yaml

@@ -1,5 +1,6 @@
 sie_id: GritLM/GritLM-7B
 hf_id: GritLM/GritLM-7B
+hf_revision: cb7f7ffb99c0c24ca2c325d798d7ef5c455d5339
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/Ihor__gliner-biomed-large-v1.0.yaml

@@ -1,5 +1,6 @@
 sie_id: Ihor/gliner-biomed-large-v1.0
 hf_id: Ihor/gliner-biomed-large-v1.0
+hf_revision: f79771210ab2111a2081ef31620fdb7342f602e9
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/Linq-AI-Research__Linq-Embed-Mistral.yaml

@@ -1,5 +1,6 @@
 sie_id: Linq-AI-Research/Linq-Embed-Mistral
 hf_id: Linq-AI-Research/Linq-Embed-Mistral
+hf_revision: 0c1a0b0589177079acc552433cad51d7c9132379
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/MoritzLaurer__ModernBERT-base-zeroshot-v2.0.yaml

@@ -1,5 +1,6 @@
 sie_id: MoritzLaurer/ModernBERT-base-zeroshot-v2.0
 hf_id: MoritzLaurer/ModernBERT-base-zeroshot-v2.0
+hf_revision: d421c4545a438fd006fb43f8b981c5d908faa1e1
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/MoritzLaurer__deberta-v3-base-zeroshot-v2.0.yaml

@@ -1,5 +1,6 @@
 sie_id: MoritzLaurer/deberta-v3-base-zeroshot-v2.0
 hf_id: MoritzLaurer/deberta-v3-base-zeroshot-v2.0
+hf_revision: 8e7e5af5983a0ddb1a5b45a38b129ab69e2258e8
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/MoritzLaurer__deberta-v3-large-zeroshot-v2.0.yaml

@@ -1,5 +1,6 @@
 sie_id: MoritzLaurer/deberta-v3-large-zeroshot-v2.0
 hf_id: MoritzLaurer/deberta-v3-large-zeroshot-v2.0
+hf_revision: cf44676c28ba7312e5c5f8f8d2c22b3e0c9cdae2
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/NovaSearch__stella_en_400M_v5.yaml

@@ -1,5 +1,6 @@
 sie_id: NovaSearch/stella_en_400M_v5
 hf_id: NovaSearch/stella_en_400M_v5
+hf_revision: ffeb2b7ee715c226d4ffe5e4619f7dbb48624c20
 inputs:
   text: true
   image: false

{sie_server-0.7.0 → sie_server-0.7.2}/models/Qwen__Qwen3-0.6B.yaml

@@ -15,11 +15,9 @@ tasks:
   # to fit the full generation benchmark pack (casehold prompts reach ~1.8k
   # tokens, gpqa ~1.3k) so the model is comparable to the rest of the fleet on
   # every task, while KV stays trivial at this size (112 KB/token → 4096 ≈
-  # 0.46 GB). The validation/co-residency harness, which packs two SGLang
-  # instances onto a single 22 GiB L4, does NOT depend on this default — it
-  # passes an explicit ``--max-seq-length``/``--context-length`` (see
-  # tools/bench_generation_matrix.py + run_generation_smoke.py) and caps itself
-  # to 1024 for that case.
+  # 0.46 GB). A deployment packing two SGLang instances onto one 22 GiB L4
+  # does NOT depend on this default: it passes an explicit
+  # ``--max-seq-length``/``--context-length`` and caps itself to 1024.
   #
   # ``context_length``, ``max_sequence_length``, and ``max_batch_tokens`` are
   # three independent knobs (per-request context, SGLang --context-length,
@@ -39,8 +37,8 @@ max_sequence_length: 4096
 # KV-cache memory math (Qwen3-0.6B, bf16):
 #   layers=28, kv_heads=8, head_dim=128, bytes_per_elem=2
 #   kv_bytes_per_token = 2 × 28 × 8 × 128 × 2 = 114,688 B ≈ 112 KB
-# At ctx=4096 a single request's KV is ~0.46 GB — negligible. The co-residency
-# harness still caps context explicitly when it has to share a card.
+# At ctx=4096 a single request's KV is ~0.46 GB — negligible. Co-resident
+# deployments can still cap context explicitly when sharing a card.
 profiles:
   default:
     max_batch_tokens: 4096
@@ -49,9 +47,8 @@ profiles:
     kv_budget_tokens: 8192
     adapter_options:
       loadtime:
-        # 0.8 leaves headroom on a 22 GiB L4 for a second SGLang instance
-        # (the validation harness co-residents worker + baseline). If the
-        # 0.6B is the only model on the card, 0.9 is fine.
+        # 0.8 leaves headroom on a 22 GiB L4 for a second SGLang instance. If
+        # the 0.6B is the only model on the card, 0.9 is fine.
         mem_fraction_static: 0.8
         served_model_name: Qwen/Qwen3-0.6B
         # Modal sandbox lacks flashinfer's JIT prerequisites; switch backends.

{sie_server-0.7.0 → sie_server-0.7.2}/models/Qwen__Qwen3-4B-Instruct-2507.yaml

@@ -38,8 +38,7 @@ tasks:
       # them on ``delta.tool_calls`` for SSE and on
       # ``message.tool_calls`` for non-streaming requests.
       tools: true
-      # Validated for code generation — MEASURED HumanEval 0.866 / MBPP 0.74 on
-      # Modal A100 (see benchmarks/generation/.../code/measured_baseline.json).
+      # Validated for code generation before enabling the code alias.
       # Backs the model="code" alias. Informational (not request-gated).
       code: true
       # Backs the model="sql" alias. NOTE: this profile runs the outlines

{sie_server-0.7.0 → sie_server-0.7.2}/models/Qwen__Qwen3.5-4B.yaml

@@ -31,8 +31,7 @@ tasks:
       # NOTE: scoped to what the production backend (``outlines``, set
       # per-profile under ``adapter_options.loadtime``) actually supports
       # on SGLang. SGLang's outlines_backend explicitly skips ebnf with
-      # ``Skip unsupported key_type='ebnf'`` — confirmed on A100 smoke
-      # 2026-05-26 (see ``benchmarks/generation/qwen35-outlines-smoke/``).
+      # ``Skip unsupported key_type='ebnf'`` — confirmed by a compatibility smoke.
       # If an operator switches a Qwen3.5 profile to ``grammar_backend:
       # xgrammar`` they should re-add ``"ebnf"`` here — xgrammar smoke on
       # the same hardware passed all three modes (regex/json_schema/ebnf).
@@ -109,9 +108,8 @@ max_sequence_length: 8192
 #     effective_max_running_requests_per_dp = 48
 #
 # ``kv_budget_tokens`` is sized at ``token_capacity / 4``, matching the
-# pattern set by Qwen3-4B-Instruct-2507. Re-calibrate via the smoke
-# tool ``tools/smoke_qwen35_4b.py`` if SGLang version or
-# ``mem_fraction_static`` changes.
+# pattern set by Qwen3-4B-Instruct-2507. Re-calibrate with an SGLang capacity
+# probe if its version or ``mem_fraction_static`` changes.
 profiles:
   # The bare model is the measured production A100-40GB shape. The
   # ``a100-40gb`` alias below preserves the resolved profile identity used by
@@ -266,8 +264,8 @@ profiles:
     max_batch_tokens: 65536
     compute_precision: bfloat16
     adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
-    # Analytical estimate — re-calibrate via ``tools/smoke_qwen35_4b.py``
-    # once H100 access is set up. Scales ~2× off the A100-40GB number
+    # Analytical estimate — re-calibrate with an SGLang capacity probe once
+    # H100 access is set up. Scales ~2× off the A100-40GB number
     # (80 GB / 40 GB), so ~420k tokens at context=32768.
     kv_budget_tokens: 421888
     adapter_options:

{sie_server-0.7.0 → sie_server-0.7.2}/models/Qwen__Qwen3.6-27B.yaml

@@ -31,19 +31,18 @@ tasks:
       grammar: ["json_schema", "regex"]
       streaming: true
       tools: true
-      # Validated on code + text-to-SQL — MEASURED on Modal H100 (bf16 no-spec,
-      # greedy+min_tokens=10; accuracy is GPU-invariant + FP8≈BF16 per ADR
-      # 0001): HumanEval 0.933, MBPP 0.81 (beats the 4B code default 0.866/0.74),
-      # Spider exec-acc 0.693. See code_sql_tools/measured_baseline.json.
+      # Validated on code + text-to-SQL — measured on H100 (bf16 no-spec,
+      # greedy+min_tokens=10). BF16 benchmark: HumanEval 0.933, MBPP 0.81
+      # (beats the 4B code default 0.866/0.74), Spider exec-acc 0.693.
       # Informational (not request-gated); surfaces the high-end code/SQL option.
       code: true
       # CAVEAT: this flag is precision-agnostic, but FP8 weight quant regresses
       # SQL ~13pts (Spider 0.71 BF16 -> 0.58 FP8, same-subset control; see the
-      # ADR + baseline). Code/tools/MC are FP8-safe. The RTX FP8 profile below is
+      # same-subset measurements). Code/tools/MC are FP8-safe. The RTX FP8 profile below is
       # for code/tools/general. Route SQL-critical traffic to the BF16 profile
       # or a BF16 bundle-qualified alias, e.g.
       # SIE_GATEWAY_MODEL_ALIASES={"sql":"<bf16-bundle>:/Qwen/Qwen3.6-27B"}
-      # (gateway resolve_model_spec_with_aliases; see ADR 0001).
+      # (gateway resolve_model_spec_with_aliases).
       sql: true
     # Qwen3.6 emits ``<think>...</think>`` reasoning by default. We
     # disable it for the OpenAI-compat path so visible output is the
@@ -217,8 +216,8 @@ profiles:
   # over H100 plus the FP8 weight saving is what makes 3/4 fit here.
   #
   # ACCURACY CONTRACT: FP8 is lossy. This profile is validated to within the
-  # Wilson 95% CI of the *BF16* baseline on all four generation tasks (see
-  # docs/adr/0001). If FP8 misses parity after bounded tuning, fall back to
+  # Wilson 95% CI of the *BF16* baseline on all four generation tasks. If FP8
+  # misses parity after bounded tuning, fall back to
   # BF16 + NEXTN 3/4 (drop the ``--quantization fp8`` pair below; the 96 GB
   # still fits the 3/4 draft in BF16). KV cache stays BF16 here — add
   # ``--kv-cache-dtype fp8_e4m3`` only if memory/throughput needs it AND