sie-server 0.4.0__tar.gz → 0.4.2__tar.gz

{sie_server-0.4.0 → sie_server-0.4.2}/Dockerfile.cpu

@@ -6,10 +6,9 @@
 #   docker buildx build --platform linux/amd64,linux/arm64 -f packages/sie_server/Dockerfile.cpu -t sie-server:cpu .
 
 ARG BUNDLE=default
-ARG SIE_DEPS_IMAGE=
 
 # =============================================================================
-# Stage 1: Dependencies (pyproject.toml only, cached across code changes)
+# Dependency image: pyproject-only cache seed
 # =============================================================================
 FROM python:3.12-slim-bookworm AS deps
 
@@ -58,9 +57,9 @@ RUN --mount=type=cache,target=/root/.cache/pip \
         -e ".[gpu-metrics]"
 
 # =============================================================================
-# Stage 2: Base - source install + shared-venv finalization (no BUNDLE)
+# Shared runtime base: source install and venv finalization
 # =============================================================================
-# Bundle-agnostic: all base-stage layers are shared across bundles of this
+# Bundle-agnostic: all base image layers are shared across bundles of this
 # platform in local BuildKit cache and in content-addressed registry layers.
 FROM deps AS base
 
@@ -108,9 +107,9 @@ RUN set -eux; \
     find /app/.venv -exec touch -h -d @0 {} + 2>/dev/null || true
 
 # =============================================================================
-# Stage 3a: Bundle-deps - bundle-specific deps install (heavy)
+# Bundle dependency builder: bundle-specific deps
 # =============================================================================
-FROM base AS bundle_deps
+FROM base AS builder
 
 ARG BUNDLE
 
@@ -151,12 +150,7 @@ RUN set -eux; \
     find /app/bundle-libs -exec touch -h -d @0 {} + 2>/dev/null || true
 
 # =============================================================================
-# Stage 3b: Builder - optional trampoline to a prebuilt base image
-# =============================================================================
-FROM ${SIE_DEPS_IMAGE:-bundle_deps} AS builder
-
-# =============================================================================
-# Stage 4: Runtime
+# Runtime image
 # =============================================================================
 FROM python:3.12-slim-bookworm AS runtime
 

{sie_server-0.4.0 → sie_server-0.4.2}/Dockerfile.cuda12

@@ -1,15 +1,14 @@
 # syntax=docker/dockerfile:1
-# SIE Server - CUDA 12.4 Image
+# SIE Server - CUDA 12 Image
 # Build from repo root:
 #   docker build -f packages/sie_server/Dockerfile.cuda12 -t sie-server:cuda12-default .
 #   docker build -f packages/sie_server/Dockerfile.cuda12 --build-arg BUNDLE=sglang -t sie-server:cuda12-sglang .
 
 ARG BUNDLE=default
 ARG UV_VERSION=0.9.28
-ARG SIE_DEPS_IMAGE=
 
 # =============================================================================
-# Stage 1: uv + standalone Python 3.12 (no deadsnakes PPA)
+# Dependency image: uv and standalone Python 3.12
 # =============================================================================
 FROM nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04 AS deps
 
@@ -60,10 +59,10 @@ RUN --mount=type=cache,target=/root/.cache/pip \
         -e ".[gpu-metrics]"
 
 # =============================================================================
-# Stage 2: Base - source install + shared-venv finalization (no BUNDLE)
+# Shared CUDA base: source install and venv finalization
 # =============================================================================
 # Everything here is bundle-agnostic, so bundle-specific builds of a given
-# platform share every base-stage layer in local BuildKit cache and in
+# platform share every base image layer in local BuildKit cache and in
 # content-addressed registry layers.
 FROM deps AS base
 
@@ -127,9 +126,9 @@ RUN set -eux; \
     find /app/.venv -exec touch -h -d @0 {} + 2>/dev/null || true
 
 # =============================================================================
-# Stage 3a: Bundle-deps - bundle-specific deps install (heavy)
+# Bundle dependency builder: bundle-specific deps
 # =============================================================================
-FROM base AS bundle_deps
+FROM base AS builder
 
 ARG BUNDLE
 
@@ -174,16 +173,19 @@ RUN set -eux; \
     find /app/bundle-libs -exec touch -h -d @0 {} + 2>/dev/null || true
 
 # =============================================================================
-# Stage 3b: Builder - optional trampoline to a prebuilt base image
+# Runtime image
 # =============================================================================
-FROM ${SIE_DEPS_IMAGE:-bundle_deps} AS builder
-
-# =============================================================================
-# Stage 4: Runtime
-# =============================================================================
-# Use base CUDA image (not devel/runtime) — PyTorch wheels bundle CUDA libs,
-# cuDNN ships inside torch. Saves ~2GB vs `runtime` variant.
-FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS runtime
+# Runtime base selection is bundle-scoped. Most CUDA bundles stay on the
+# smaller CUDA base runtime; SGLang-family bundles need the devel toolkit
+# because flashinfer/tvm_ffi perform runtime JIT through nvcc on first decode.
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS runtime-default
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS runtime-transformers5
+FROM nvidia/cuda:12.9.1-cudnn-devel-ubuntu22.04 AS runtime-sglang
+ENV CUDA_HOME=/usr/local/cuda \
+    LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
+FROM runtime-sglang AS runtime-sglang-embedding
+
+FROM runtime-${BUNDLE} AS runtime
 
 ENV DEBIAN_FRONTEND=noninteractive
 

{sie_server-0.4.0 → sie_server-0.4.2}/PKG-INFO

@@ -1,10 +1,11 @@
 Metadata-Version: 2.4
 Name: sie-server
-Version: 0.4.0
+Version: 0.4.2
 Summary: Search Inference Engine - GPU inference server for search workloads
 License: Apache-2.0
 License-File: LICENSE
 Requires-Python: <3.13,>=3.12
+Requires-Dist: blake3<1,>=0.4
 Requires-Dist: docling<3,>=2
 Requires-Dist: einops<1,>=0.8
 Requires-Dist: fastapi<1,>=0.115
@@ -17,7 +18,6 @@ Requires-Dist: loguru<1,>=0.7
 Requires-Dist: msgpack-numpy<1,>=0.4
 Requires-Dist: msgpack<2,>=1.1
 Requires-Dist: msgspec>=0.20.0
-Requires-Dist: nats-py<3,>=2.9
 Requires-Dist: numpy<3,>=2
 Requires-Dist: open-clip-torch>=2.24
 Requires-Dist: opencv-python-headless<5,>=4
@@ -26,7 +26,7 @@ Requires-Dist: opentelemetry-exporter-otlp<2,>=1.28
 Requires-Dist: opentelemetry-instrumentation-fastapi<1,>=0.49b0
 Requires-Dist: opentelemetry-sdk<2,>=1.28
 Requires-Dist: packaging<25,>=24
-Requires-Dist: pillow<12,>=11
+Requires-Dist: pillow>=12.2.0
 Requires-Dist: prometheus-client<1,>=0.21
 Requires-Dist: pydantic-settings<3,>=2.6
 Requires-Dist: pydantic<3,>=2.9

{sie_server-0.4.0 → sie_server-0.4.2}/README.md

@@ -66,7 +66,7 @@ auto-retries; see `packages/sie_sdk/README.md` for client-side controls.
 
 | Env var | Default | Effect |
 |--|--|--|
-| `SIE_GRAMMAR_PREFLIGHT_DEBUG` | unset (off) | Enables the legacy worker-side Outlines preflight compile before each structured-output request. Off by default per ADR-0002 — SGLang is the production grammar authority. Use for diagnosing schema-rejection problems or slow compiles in a controlled environment; not recommended for production traffic. |
+| `SIE_GRAMMAR_PREFLIGHT_DEBUG` | unset (off) | Enables the legacy worker-side Outlines preflight compile before each structured-output request. Off by default because SGLang is the production grammar authority. Use for diagnosing schema-rejection problems or slow compiles in a controlled environment; not recommended for production traffic. |
 
 For nested settings (any field with `__`), the env-var format is
 `SIE_<TOP>__<NESTED>=value`. The complete schema is in

{sie_server-0.4.0 → sie_server-0.4.2}/bundles/default.yaml

@@ -43,6 +43,7 @@ adapters:
 - sie_server.adapters.florence2
 - sie_server.adapters.docling
 - sie_server.adapters.paddleocr_vl
+- sie_server.adapters.mineru_vl
 deps:
   # Most flash adapters; sentence_transformer needs >=4.57
   transformers: '>=4.57,<5'
@@ -78,5 +79,5 @@ deps:
   docling: '>=2,<3'
   # Flash Attention 2 — CUDA only, prebuilt wheel
   flash-attn:
-    url: https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.0/flash_attn-2.7.4+cu128torch2.9-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
+    url: https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.7.11/flash_attn-2.7.4+cu129torch2.9-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
     marker: sys_platform == 'linux'

{sie_server-0.4.0 → sie_server-0.4.2}/bundles/sglang-embedding.yaml

@@ -12,7 +12,7 @@ deps:
   # pip resolution drift on environments that already had a different
   # ``outlines`` installed. Asserted in ``tests/test_bundles.py`` so a
   # future drift fails fast.
-  sglang: '==0.5.10'
+  sglang: '==0.5.10.post1'
   xgrammar: '==0.1.32'
   outlines: '==0.1.11'
   llguidance: '>=0.7.11,<0.8.0'

{sie_server-0.4.0 → sie_server-0.4.2}/bundles/sglang.yaml

@@ -6,29 +6,29 @@ deps:
   # SGLang 0.5.6+ includes all deps at base level (fixed from 0.4.x extras bug)
   # See: https://github.com/sgl-project/sglang/issues/4869
   #
-  # Qwen3.5-4B compatibility — M4 req2 Proj 5:
+  # Qwen3.5-4B + Qwen3.6-27B compatibility:
   #
-  # ``sglang==0.5.10`` is the canonical target for Qwen3.5-4B on the
-  # current L4 / A100-40GB / H100 fleet. Audited against
-  # ``python/pyproject.toml@v0.5.10`` upstream (see
-  # ``product/plans/qwen35-sglang-mtp-structured-outputs-findings.md``):
+  # ``sglang==0.5.10.post1`` is the canonical target for the Qwen3.x hybrid
+  # Gated-DeltaNet + Gated-Attention family on the current
+  # L4 / A100-40GB / H100 fleet. Qwen3.6-27B uses the same ``qwen3_5``
+  # model class shipped in 0.5.10 — the architecture (64 layers, hybrid
+  # Gated DeltaNet + Gated Attention, MTP/NEXTN) is identical, only the
+  # parameter count differs.
   #
-  #   * ships the ``qwen3_5`` model class (``models/qwen3_5.py``, 1724 LOC)
-  #   * grammar backends: ``xgrammar`` / ``outlines`` / ``llguidance`` / ``none``
-  #   * ``sglang-kernel==0.4.1`` wheel covers SM_80 / SM_89 / SM_90 / SM_100
-  #     via gencode (``CMakeLists.txt``: ``ENABLE_BELOW_SM90=ON`` default).
-  #     The runtime loader (``sgl_kernel/load_utils.py``) maps
-  #     compute_capability != 90 → ``sm100/`` subdir, which holds the
-  #     SM_80 / SM_89 / SM_100 build (precise math). H100 (CC=90) gets
-  #     the ``sm90/`` fast-math build.
-  #   * torch==2.9.1 (CUDA 12.8/12.9 wheels); ``cuda-python==12.9``.
-  #     **Not** CUDA 13 — that's an SGLang-main-only path which only
-  #     became relevant when looking at the dev branch.
+  # SGLang 0.5.10 was evaluated against CUDA 12.9 + Qwen3.6-27B on Modal
+  # H100 (2026-05-27): server boots, loads weights, but the bundled
+  # ``sglang/jit_kernel/csrc/elementwise/activation.cuh`` has a C++
+  # template bug (``select_kernel<true>(type)`` is parsed as a class-
+  # template substitution, not a function-template call) that the
+  # stricter ``nvcc`` shipped with CUDA 12.9 rejects at first activation.
+  # 0.5.11 is also dev-only on the sglang docs wheel index — not on
+  # PyPI. Park the 0.5.11 bump until upstream cuts a stable release with
+  # the JIT header fixed; 0.5.10.post1 covers Qwen3.6-27B today.
   #
-  # Compat note: 0.5.12 wheel observed shipping only ``sm100/`` (no SM_80
-  # cubin entry inside) — out of scope; we stay on 0.5.10 until upstream
-  # ships multi-arch binaries again.
-  sglang: '==0.5.10'
+  #   * grammar backends: ``xgrammar`` / ``outlines`` / ``llguidance`` / ``none``
+  #   * ``sgl_kernel`` covers SM_80 / SM_89 / SM_90 / SM_100 via gencode.
+  #   * torch==2.9.1 (CUDA 12.9 wheels); ``cuda-python==12.9``.
+  sglang: '==0.5.10.post1'
   #
   # Grammar backend deps — pinned to exactly what SGLang 0.5.10 imports
   # internally to prevent silent pip resolution drift.
@@ -39,7 +39,7 @@ deps:
   # propagation (PR #20467). Kept available as the fallback backend.
   xgrammar: '==0.1.32'
   #
-  # ``outlines==0.1.11`` is a hard transitive dep of sglang==0.5.10. We
+  # ``outlines==0.1.11`` is a hard transitive dep of sglang==0.5.10.post1. We
   # declare it explicitly at bundle level so the surface is visible.
   # ``outlines-core`` (a separate package) is a transitive of outlines
   # and intentionally NOT pinned here — pinning ``outlines-core`` directly

sie_server-0.4.2/models/Marqo__marqo-fashionSigLIP.yaml

@@ -0,0 +1,28 @@
+sie_id: Marqo/marqo-fashionSigLIP
+hf_id: Marqo/marqo-fashionSigLIP
+inputs:
+  text: true
+  image: true
+  audio: false
+  video: false
+tasks:
+  encode:
+    dense:
+      dim: 768
+    sparse: null
+    multivector: null
+  score: null
+  extract: null
+max_sequence_length: 64
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: float16
+    adapter_path: sie_server.adapters.siglip:SiglipAdapter
+    adapter_options:
+      loadtime:
+        backend: open_clip
+        open_clip_model_id: hf-hub:Marqo/marqo-fashionSigLIP
+        dense_dim: 768
+      runtime:
+        normalize: true

{sie_server-0.4.0 → sie_server-0.4.2}/models/Qwen__Qwen3-0.6B.yaml

@@ -6,40 +6,39 @@ inputs:
   audio: false
   video: false
 tasks:
-  # Tiny generation model used for fast e2e iteration on Modal L4 (walking-skeleton
-  # viability bench + validation-harness SIE-only smoke). Loads in ~30s. Quality is
-  # bad — this is a transport benchmark target, not a production model.
+  # Small, fast generation model — a viable PROD pick for simple/short-prompt
+  # tasks (cheap + high throughput; weaker on long-context reasoning). Loads in
+  # ~30s. Also doubles as the transport/walking-skeleton benchmark target.
   #
-  # Context / batch sizes are deliberately small (1024 vs the headroom an
-  # L4 could nominally support) so the validation harness can co-resident
-  # the worker's SGLang with a second SGLang for the baseline phase on
-  # a single 22 GiB card without OOMing. Raise these if you ever run 0.6B
-  # alone on a card it doesn't have to share.
+  # ``context_length`` is the standalone PROD serving value (4096): big enough
+  # 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.
   #
-  # Note on the three 1024s below: `context_length`, `max_sequence_length`,
-  # and `max_batch_tokens` are NOT redundant — they're three independent
-  # knobs (per-request context, SGLang --context-length, batcher cost
-  # budget) that just happen to collide here because the model is tiny.
-  # See sibling Qwen__Qwen3-4B-Instruct-2507.yaml for the canonical
-  # non-collapsed shape.
+  # ``context_length``, ``max_sequence_length``, and ``max_batch_tokens`` are
+  # three independent knobs (per-request context, SGLang --context-length,
+  # batcher cost budget); see sibling Qwen__Qwen3-4B-Instruct-2507.yaml.
   generate:
-    context_length: 1024
+    context_length: 4096
     max_output_tokens: 1024
     capabilities:
       grammar: []
       streaming: true
       tools: false
-max_sequence_length: 1024
+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
-# The 0.6B is a transport benchmark target — context_length is held at
-# 1024 deliberately (see header comment) so the validation harness can
-# co-resident two SGLang instances on an L4. KV budgets per profile
-# scale with the deployment scenario rather than the GPU ceiling.
+# 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.
 profiles:
   default:
-    max_batch_tokens: 1024
+    max_batch_tokens: 4096
     compute_precision: bfloat16
     adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
     kv_budget_tokens: 8192

{sie_server-0.4.0 → sie_server-0.4.2}/models/Qwen__Qwen3-4B-Instruct-2507.yaml

@@ -66,8 +66,7 @@ max_sequence_length: 32768
 # kv_budget_tokens is set to ~40% of theoretical max, matching the L4
 # baseline ratio (32768/90000 ≈ 36%). The headroom absorbs batch growth,
 # speculative side-cell, grammar/Outlines compile arena, fragmentation.
-# Final empirical validation (concurrency-16 OOM-boundary sweep) is
-# tracked in product/plans/m4-req2-gpu-runbook.md §"#16/#19".
+# Final empirical validation should use concurrency and OOM-boundary sweeps.
 profiles:
   default:
     # max_batch_tokens is a generic engine knob; generation does not batch
@@ -93,7 +92,7 @@ profiles:
           top_p: 0.9
         stop_tokens:
           - "<|im_end|>"
-  # M5 audit #16/#19: analytical defaults for a100-40gb / h100. Production
+  # Analytical defaults for a100-40gb / h100. Production
   # capacity also grows: with 2-4× the KV budget the context window can be
   # widened proportionally so longer-context workloads (RAG with large
   # retrieved passages) fit comfortably. ``max_output_tokens`` doubles