sie-server 0.3.4__tar.gz → 0.4.1__tar.gz

{sie_server-0.3.4 → sie_server-0.4.1}/Dockerfile.cpu

@@ -63,6 +63,13 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # platform in local BuildKit cache and in content-addressed registry layers.
 FROM deps AS base
 
+# Source-layer cache key — see Dockerfile.cuda12 for the full rationale.
+# Tie source layers to a per-commit revision arg so a rebuild can't ship
+# stale code via a reused source-COPY layer. Dependency layers in the
+# ``deps`` stage above stay cached.
+ARG SIE_SRC_REV=dev
+RUN echo "sie source revision: ${SIE_SRC_REV}"
+
 COPY packages/sie_sdk/src /tmp/sie_sdk/src
 COPY packages/sie_server/src src/
 COPY packages/sie_server/bundles bundles/
@@ -158,13 +165,25 @@ ENV DEBIAN_FRONTEND=noninteractive \
 # Only the shared libs torch + pillow + rtree actually dlopen at runtime.
 # libspatialindex-c6: rtree (docling dep) dlopens libspatialindex_c.so; the
 # rtree==1.4.1 wheel only bundles the C++ core, not the C wrapper.
+# libgl1 libglib2.0-0 libice6 libsm6 libx11-6 libxcb1 libxext6: docling-ibm-models'
+# TableFormer imports cv2 during DoclingAdapter.load() pre-warm; the opencv-python
+# wheel unconditionally dlopens an X11 + libGL + glib chain at import even in
+# headless usage. Without these, every docling extract crashes with
+# "ImportError: libxcb.so.1: cannot open shared object file" (issue #1028).
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
+    libgl1 \
+    libglib2.0-0 \
     libgomp1 \
+    libice6 \
     libjpeg62-turbo \
     libpng16-16 \
-    libspatialindex-c6
+    libsm6 \
+    libspatialindex-c6 \
+    libx11-6 \
+    libxcb1 \
+    libxext6
 
 RUN groupadd -g 1000 sie && useradd -u 1000 -g sie -m sie
 
@@ -176,8 +195,10 @@ WORKDIR /app
 # no /etc/passwd visible (the sie user exists in the runtime FS but --link
 # layers are created in isolation).
 COPY --link --from=base --chown=1000:1000 /app/.venv /app/.venv
-COPY --link --from=base --chown=1000:1000 /app/src /app/src
-COPY --link --from=base --chown=1000:1000 /tmp/sie_sdk/src /tmp/sie_sdk/src
+# Source trees WITHOUT --link — see Dockerfile.cuda12 (linked cross-stage
+# copies didn't reliably invalidate on source change).
+COPY --from=base --chown=1000:1000 /app/src /app/src
+COPY --from=base --chown=1000:1000 /tmp/sie_sdk/src /tmp/sie_sdk/src
 COPY --link --from=base --chown=1000:1000 /app/models /app/models
 COPY --link --from=base --chown=1000:1000 /app/bundles /app/bundles
 # Bundle-specific extras — last layer so shared layers above stay cached.

{sie_server-0.3.4 → sie_server-0.4.1}/Dockerfile.cuda12

@@ -66,6 +66,18 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # content-addressed registry layers.
 FROM deps AS base
 
+# Source-layer cache key. A rebuild once shipped STALE code: buildx
+# reused a cached source-COPY layer even though the .py files had
+# changed, so the demo had to overlay patched files by hand. Tie the
+# source layers to an explicit revision arg the CI passes per commit
+# (``--build-arg SIE_SRC_REV=$(git rev-parse --short HEAD)``) so any
+# commit forces these layers — and the editable reinstall below — to
+# rebuild. The expensive dependency install lives in the ``deps`` stage
+# ABOVE this line, so it stays cached. Bundles of the same commit share
+# the same SIE_SRC_REV, so cross-bundle layer dedup is preserved.
+ARG SIE_SRC_REV=dev
+RUN echo "sie source revision: ${SIE_SRC_REV}"
+
 COPY packages/sie_sdk/src /tmp/sie_sdk/src
 COPY packages/sie_server/src src/
 COPY packages/sie_server/bundles bundles/
@@ -175,15 +187,27 @@ ENV DEBIAN_FRONTEND=noninteractive
 # libgomp1: torch OpenMP runtime.
 # libspatialindex-c6: rtree (docling dep) dlopens libspatialindex_c.so; the
 #   rtree==1.4.1 wheel only bundles the C++ core, not the C wrapper.
+# libgl1 libglib2.0-0 libice6 libsm6 libx11-6 libxcb1 libxext6: docling-ibm-models'
+#   TableFormer imports cv2 during DoclingAdapter.load() pre-warm; the
+#   opencv-python wheel unconditionally dlopens an X11 + libGL + glib chain at
+#   import even in headless usage. Without these, every docling extract crashes
+#   with "ImportError: libxcb.so.1: cannot open shared object file" (issue #1028).
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
     ca-certificates \
     gcc \
     libc6-dev \
+    libgl1 \
+    libglib2.0-0 \
     libgomp1 \
+    libice6 \
     libnuma1 \
-    libspatialindex-c6
+    libsm6 \
+    libspatialindex-c6 \
+    libx11-6 \
+    libxcb1 \
+    libxext6
 
 RUN groupadd -g 1000 sie && useradd -u 1000 -g sie -m sie
 
@@ -211,8 +235,13 @@ RUN set -e; \
 # (the sie user is added in the runtime stage filesystem but --link layers
 # are created in isolation from the destination stage state).
 COPY --link --from=base --chown=1000:1000 /app/.venv /app/.venv
-COPY --link --from=base --chown=1000:1000 /app/src /app/src
-COPY --link --from=base --chown=1000:1000 /tmp/sie_sdk/src /tmp/sie_sdk/src
+# Source trees are copied WITHOUT --link. ``COPY --link --from=<stage>``
+# layers are cached on a digest that, in the buildx versions this image
+# was built with, did not reliably invalidate when the upstream source
+# changed — the stale-code bug above. These trees are small, so dropping
+# --link costs negligible dedup while guaranteeing edited code ships.
+COPY --from=base --chown=1000:1000 /app/src /app/src
+COPY --from=base --chown=1000:1000 /tmp/sie_sdk/src /tmp/sie_sdk/src
 COPY --link --from=base --chown=1000:1000 /app/models /app/models
 COPY --link --from=base --chown=1000:1000 /app/bundles /app/bundles
 # Bundle-specific extras — last layer so shared layers above stay cached.

{sie_server-0.3.4 → sie_server-0.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sie-server
-Version: 0.3.4
+Version: 0.4.1
 Summary: Search Inference Engine - GPU inference server for search workloads
 License: Apache-2.0
 License-File: LICENSE

{sie_server-0.3.4 → sie_server-0.4.1}/README.md

@@ -62,6 +62,12 @@ auto-retries; see `packages/sie_sdk/README.md` for client-side controls.
 | `SIE_DEFAULT_COMPUTE_PRECISION` | `float16` | One of `float16`, `bfloat16`, `float32`. |
 | `SIE_ATTENTION_BACKEND` | `auto` | One of `auto`, `flash_attention_2`, `sdpa`, `eager`. |
 
+### Diagnostics
+
+| 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. |
+
 For nested settings (any field with `__`), the env-var format is
 `SIE_<TOP>__<NESTED>=value`. The complete schema is in
 `packages/sie_server/src/sie_server/config/engine.py`.

{sie_server-0.3.4 → sie_server-0.4.1}/bundles/default.yaml

@@ -78,5 +78,5 @@ 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-0.4.1/bundles/sglang-embedding.yaml

@@ -0,0 +1,18 @@
+name: sglang-embedding
+priority: 21
+adapters:
+- sie_server.adapters.sglang.embedding
+deps:
+  # Lockstep with ``bundles/sglang.yaml``. The dependency stack is
+  # identical (sglang's grammar backends are unavoidably pulled in even
+  # on embedding-only deployments), and the bundle split exists only for
+  # worker pool isolation. The previous revision pinned ``outlines-core``
+  # (a transitive of ``outlines``) instead of the actual top-level deps
+  # the generation bundle pins, which was both wrong and produced silent
+  # 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.post1'
+  xgrammar: '==0.1.32'
+  outlines: '==0.1.11'
+  llguidance: '>=0.7.11,<0.8.0'

sie_server-0.4.1/bundles/sglang.yaml

@@ -0,0 +1,66 @@
+name: sglang
+priority: 20
+adapters:
+- sie_server.adapters.sglang.generation
+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 + Qwen3.6-27B compatibility:
+  #
+  # ``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.
+  #
+  # 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.
+  #
+  #   * 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.
+  #
+  # XGrammar is SGLang 0.5.10's other supported grammar backend; it pins
+  # ``xgrammar==0.1.32`` exactly. That release also brought the
+  # structured-output VRAM-leak fix (PR #20697) and grammar-error
+  # 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.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
+  # was a bug in the previous bundle revision (it does not give us any
+  # grammar functionality on its own).
+  #
+  # Outlines IS the active grammar backend for Qwen3.5 (partner
+  # requirement). Earlier revisions said "do not switch Qwen3.5 to
+  # outlines" because the worker-side ``compile_outlines`` preflight
+  # crashed with ``'TokenizersBackend' object has no attribute
+  # 'vocabulary'``: it passed the raw transformers==5.3.0 tokenizer
+  # (now a ``TokenizersBackend``) to Outlines' processor factories, which
+  # require an Outlines ``Tokenizer`` adapter exposing ``.vocabulary``.
+  # ``compile_outlines`` now wraps the tokenizer in Outlines'
+  # ``TransformerTokenizer`` first (the same wrap SGLang's
+  # ``OutlinesGrammarBackend`` does internally), so the mismatch is gone
+  # for json_schema/regex; ebnf is forwarded straight to SGLang. See
+  # ``processors/grammar_compile.py`` ("Tokenizer adapter").
+  outlines: '==0.1.11'
+  #
+  # llguidance is the third grammar backend (regex / json_schema / ebnf).
+  # Pinned to SGLang 0.5.10's compatible range. Kept available as the
+  # fallback if a future model regresses on xgrammar.
+  llguidance: '>=0.7.11,<0.8.0'

{sie_server-0.3.4 → sie_server-0.4.1}/models/Alibaba-NLP__gte-Qwen2-7B-instruct.yaml

@@ -18,7 +18,7 @@ profiles:
   default:
     max_batch_tokens: 16384
     compute_precision: bfloat16
-    adapter_path: sie_server.adapters.sglang:SGLangEmbeddingAdapter
+    adapter_path: sie_server.adapters.sglang.embedding:SGLangEmbeddingAdapter
     adapter_options:
       loadtime:
         mem_fraction_static: 0.85

{sie_server-0.3.4 → sie_server-0.4.1}/models/Linq-AI-Research__Linq-Embed-Mistral.yaml

@@ -18,7 +18,7 @@ profiles:
   default:
     max_batch_tokens: 8192
     compute_precision: bfloat16
-    adapter_path: sie_server.adapters.sglang:SGLangEmbeddingAdapter
+    adapter_path: sie_server.adapters.sglang.embedding:SGLangEmbeddingAdapter
     adapter_options:
       loadtime:
         mem_fraction_static: 0.85

sie_server-0.4.1/models/Qwen__Qwen3-0.6B.yaml

@@ -0,0 +1,119 @@
+sie_id: Qwen/Qwen3-0.6B
+hf_id: Qwen/Qwen3-0.6B
+inputs:
+  text: true
+  image: false
+  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.
+  #
+  # 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.
+  #
+  # 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.
+  generate:
+    context_length: 1024
+    max_output_tokens: 1024
+    capabilities:
+      grammar: []
+      streaming: true
+      tools: false
+max_sequence_length: 1024
+# 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.
+profiles:
+  default:
+    max_batch_tokens: 1024
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
+    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.
+        mem_fraction_static: 0.8
+        served_model_name: Qwen/Qwen3-0.6B
+        # Modal sandbox lacks flashinfer's JIT prerequisites; switch backends.
+        disable_cuda_graph: true
+        attention_backend: triton
+      runtime:
+        first_chunk_timeout_s: 30
+        inter_chunk_timeout_s: 10
+        # Aligned with the rest of the generate model fleet (300s).
+        # The previous 132s was an unexplained magic number that
+        # diverged from every other generate config; bumping to the
+        # fleet default keeps long-completion requests from hitting a
+        # premature overall-timeout on the 0.6B model.
+        overall_timeout_s: 300
+        default_sampling:
+          temperature: 0.0
+          top_p: 1.0
+  # Dedicated 0.6B deployments on a100/h100 don't co-resident a baseline,
+  # so mem_fraction_static returns to the standard 0.85 and the KV budget
+  # scales with the larger GPU. kv_budget_tokens stays well below the
+  # theoretical ceiling because the 0.6B's *context_length* (1024) caps
+  # per-request KV consumption — the budget really just sets the upper
+  # bound on concurrent in-flight sequences.
+  a100-40gb:
+    max_batch_tokens: 4096
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
+    kv_budget_tokens: 32768
+    adapter_options:
+      loadtime:
+        mem_fraction_static: 0.85
+        served_model_name: Qwen/Qwen3-0.6B
+        disable_cuda_graph: true
+        attention_backend: triton
+      runtime:
+        first_chunk_timeout_s: 30
+        inter_chunk_timeout_s: 10
+        # Aligned with the rest of the generate model fleet (300s).
+        # The previous 132s was an unexplained magic number that
+        # diverged from every other generate config; bumping to the
+        # fleet default keeps long-completion requests from hitting a
+        # premature overall-timeout on the 0.6B model.
+        overall_timeout_s: 300
+        default_sampling:
+          temperature: 0.0
+          top_p: 1.0
+  h100:
+    max_batch_tokens: 8192
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
+    kv_budget_tokens: 65536
+    adapter_options:
+      loadtime:
+        mem_fraction_static: 0.85
+        served_model_name: Qwen/Qwen3-0.6B
+        disable_cuda_graph: true
+        attention_backend: triton
+      runtime:
+        first_chunk_timeout_s: 30
+        inter_chunk_timeout_s: 10
+        # Aligned with the rest of the generate model fleet (300s).
+        # The previous 132s was an unexplained magic number that
+        # diverged from every other generate config; bumping to the
+        # fleet default keeps long-completion requests from hitting a
+        # premature overall-timeout on the 0.6B model.
+        overall_timeout_s: 300
+        default_sampling:
+          temperature: 0.0
+          top_p: 1.0

sie_server-0.4.1/models/Qwen__Qwen3-4B-Instruct-2507.yaml

@@ -0,0 +1,152 @@
+sie_id: Qwen/Qwen3-4B-Instruct-2507
+hf_id: Qwen/Qwen3-4B-Instruct-2507
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  # kv_budget_tokens now lives on profiles (below). The KV calibration
+  # follow-up publishes the tuned value; the placeholder here is conservative
+  # and assumes Qwen3-4B's ~150 KB/token KV footprint on an L4 (24 GB,
+  # mem_fraction_static=0.85).
+  generate:
+    context_length: 32768
+    max_output_tokens: 4096
+    capabilities:
+      # Outlines-backed JSON Schema, regex, and EBNF grammars are
+      # all supported by the SGLang adapter (Outlines and XGrammar
+      # both accept EBNF natively). The gateway gates requests on
+      # this exact list — adding a new ``grammar.kind`` variant
+      # requires both the gateway parser and this list to be updated.
+      grammar: ["json_schema", "regex", "ebnf"]
+      streaming: true
+      # Qwen3-4B-Instruct's chat template emits OpenAI-compatible
+      # ``<tool_call>{...}</tool_call>`` blocks when ``tools`` is
+      # present in the rendered messages; the worker's
+      # ``parse_tool_call_stream`` consumes those blocks and surfaces
+      # them on ``delta.tool_calls`` for SSE and on
+      # ``message.tool_calls`` for non-streaming requests.
+      tools: true
+    # Forwarded verbatim to ``tokenizer.apply_chat_template(**kwargs)`` when
+    # the worker renders an OpenAI-shaped ``messages`` request.
+    # Qwen3's chat template emits a ``<think>``/``</think>`` reasoning block
+    # unless this flag suppresses it.
+    chat_template_kwargs:
+      enable_thinking: false
+    # Schemas/regexes the worker pre-compiles at model load so the first
+    # request hitting them skips the Outlines compile (cold TTFT win).
+    # Failures here log + bump ``sie_worker_grammar_prewarm_total{outcome="failed"}``
+    # without blocking model load — add entries only for shapes you
+    # know are hot.
+    prewarm_grammars:
+      # Bare pattern, NOT anchored: this model uses the default Outlines
+      # grammar backend, and Outlines regexes are implicitly anchored —
+      # its FSM engine (interegular) rejects ``^``/``$`` with
+      # ``Unsupported``, which crashes SGLang's scheduler. Use ``(yes|no)``.
+      - name: yes_no
+        kind: regex
+        value: "(yes|no)"
+      - name: short_answer
+        kind: json_schema
+        value:
+          type: object
+          properties:
+            answer:
+              type: string
+          required: [answer]
+max_sequence_length: 32768
+# KV-cache memory math (Qwen3-4B-Instruct-2507, bf16):
+#   layers=36, kv_heads=8, head_dim=128, bytes_per_elem=2
+#   kv_bytes_per_token = 2 (k+v) × 36 × 8 × 128 × 2 = 147,456 B ≈ 144 KB
+# Theoretical max KV tokens per GPU (assuming ~8 GB weights, mem_fraction_static=0.85):
+#   l4         (24 GB):  (24 × 0.85 − 8) GB / 144 KB ≈   90,000 tokens
+#   a100-40gb  (40 GB):  (40 × 0.85 − 8) GB / 144 KB ≈  189,000 tokens
+#   h100       (80 GB):  (80 × 0.85 − 8) GB / 144 KB ≈  437,000 tokens
+# 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".
+profiles:
+  default:
+    # max_batch_tokens is a generic engine knob; generation does not batch
+    # at the SIE layer (SGLang batches internally) but the validator
+    # requires the field to be set.
+    max_batch_tokens: 16384
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
+    # L4 baseline — empirically gated by the speculative-decoding + calibration work.
+    kv_budget_tokens: 32768
+    adapter_options:
+      loadtime:
+        mem_fraction_static: 0.85
+        served_model_name: Qwen/Qwen3-4B-Instruct-2507
+        # speculative decoding (MTP/EAGLE/NGRAM) intentionally absent;
+        # week-1 validation decides whether to promote a side-cell. See §4.9.
+      runtime:
+        first_chunk_timeout_s: 30
+        inter_chunk_timeout_s: 10
+        overall_timeout_s: 300
+        default_sampling:
+          temperature: 0.7
+          top_p: 0.9
+        stop_tokens:
+          - "<|im_end|>"
+  # M5 audit #16/#19: 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
+  # to 8192/16384 respectively — beyond that, latency hurts more than
+  # quality helps for instruction-style chat traffic.
+  a100-40gb:
+    max_batch_tokens: 32768
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
+    # Empirically calibrated on Modal A100-SXM4-40GB
+    # (sglang 0.5.9, mem_fraction_static=0.85, context_length=32768).
+    # SGLang's /server_info reports:
+    #   weight=7.71 GB, kvcache=25.42 GB, graph=0.18 GB,
+    #   token_capacity=185,081 tokens
+    # ``kv_budget_tokens`` sized for 4 concurrent admissions:
+    #   185,081 / 4 = 46,270 → round down to 45,056 for headroom.
+    # Re-calibrate if SGLang version or mem_fraction_static changes.
+    kv_budget_tokens: 45056
+    adapter_options:
+      loadtime:
+        mem_fraction_static: 0.85
+        served_model_name: Qwen/Qwen3-4B-Instruct-2507
+      runtime:
+        first_chunk_timeout_s: 30
+        inter_chunk_timeout_s: 10
+        overall_timeout_s: 300
+        default_sampling:
+          temperature: 0.7
+          top_p: 0.9
+        stop_tokens:
+          - "<|im_end|>"
+  h100:
+    max_batch_tokens: 65536
+    compute_precision: bfloat16
+    adapter_path: sie_server.adapters.sglang.generation:SGLangGenerationAdapter
+    # Empirically calibrated on Modal H100 80GB HBM3
+    # (sglang 0.5.9, mem_fraction_static=0.85, context_length=32768).
+    # SGLang's /server_info reports:
+    #   weight=7.71 GB, kvcache=59.0 GB, graph=0.43 GB,
+    #   token_capacity=429,645 tokens
+    # ``kv_budget_tokens`` sized for 4 concurrent admissions:
+    #   429,645 / 4 = 107,411 → round down to 106,496 for headroom.
+    kv_budget_tokens: 106496
+    adapter_options:
+      loadtime:
+        mem_fraction_static: 0.85
+        served_model_name: Qwen/Qwen3-4B-Instruct-2507
+      runtime:
+        first_chunk_timeout_s: 30
+        inter_chunk_timeout_s: 10
+        overall_timeout_s: 300
+        default_sampling:
+          temperature: 0.7
+          top_p: 0.9
+        stop_tokens:
+          - "<|im_end|>"

{sie_server-0.3.4 → sie_server-0.4.1}/models/Qwen__Qwen3-Embedding-4B.yaml

@@ -18,7 +18,7 @@ profiles:
   default:
     max_batch_tokens: 16384
     compute_precision: bfloat16
-    adapter_path: sie_server.adapters.sglang:SGLangEmbeddingAdapter
+    adapter_path: sie_server.adapters.sglang.embedding:SGLangEmbeddingAdapter
     adapter_options:
       loadtime:
         mem_fraction_static: 0.85