kanadaj · kanadaj · Sep 14, 2026 · Sep 14, 2026 · Sep 14, 2026 · Sep 14, 2026
diff --git a/Dockerfile.hicache-wip b/Dockerfile.hicache-wip
@@ -0,0 +1,14 @@
+# EXPERIMENTAL: opt-in HiCache profile, isolated from the default images.
+# Configuration-specific qualification is recorded in docs/hicache-wip.md.
+FROM docker.io/kanadaj/sglang-qwen38fn-sm120-turbo@sha256:f2859d1ccf824a5295088cf578eba89b0f3eeefff6ae7679c3f5d64af0689458
+ARG SOURCE_REVISION
+LABEL org.opencontainers.image.source="https://github.com/kanadaj/sglang" \
+      org.opencontainers.image.revision="${SOURCE_REVISION}" \
+      org.opencontainers.image.description="Opt-in Qwen HiCache state-transfer profile; qualify the deployment configuration"
+COPY patches/0020-hicache-ple-state.patch patches/0021-hicache-file-integrity.patch patches/0022-hicache-qsa-sidecar.patch patches/0023-qsa-sparse-gather-memory-safety.patch patches/0024-router-pdl-bias-order.patch patches/0025-hicache-load-order.patch patches/0026-qsa-short-extend-bounds.patch patches/series.hicache-wip /opt/qwen-hicache-wip/patches/
+COPY provenance/invalid-token-failure-runtime-files.json provenance/hicache-wip.json /opt/qwen-hicache-wip/provenance/
+COPY scripts/verify_hicache_wip.py /opt/qwen-hicache-wip/scripts/verify_hicache_wip.py
+RUN python3 -B /opt/qwen-hicache-wip/scripts/verify_hicache_wip.py \
+      --tree /sgl-workspace/sglang --apply
+ENTRYPOINT ["python3", "-m", "sglang.launch_server"]
+CMD ["--help"]
diff --git a/Dockerfile.invalid-token-failure b/Dockerfile.invalid-token-failure
@@ -0,0 +1,19 @@
+# Cumulative CPU-only overlay: patches 0015 through 0019.
+# All model paths, serving arguments, and runtime settings remain external.
+FROM docker.io/kanadaj/sglang-qwen38fn-sm120-turbo@sha256:872a2bda228e39aa9c1af729b47cc28f7862e7859e448f1a8868b85a4051f404
+ARG SOURCE_REVISION
+LABEL org.opencontainers.image.source="https://github.com/kanadaj/sglang" \
+      org.opencontainers.image.revision="${SOURCE_REVISION}"
+COPY runtime.invalid-token-failure/python/sglang/srt/entrypoints/openai/serving_chat.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_chat.py
+COPY runtime.invalid-token-failure/python/sglang/srt/entrypoints/openai/serving_completions.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_completions.py
+COPY runtime/python/sglang/srt/entrypoints/openai/protocol.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/protocol.py
+COPY runtime.invalid-token-failure/python/sglang/srt/entrypoints/openai/serving_responses.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_responses.py
+COPY runtime/python/sglang/srt/entrypoints/openai/responses_compat.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/responses_compat.py
+COPY runtime/python/sglang/srt/function_call/qwen3_coder_detector.py /sgl-workspace/sglang/python/sglang/srt/function_call/qwen3_coder_detector.py
+COPY runtime/python/sglang/srt/multimodal/processors/qwen_vl.py /sgl-workspace/sglang/python/sglang/srt/multimodal/processors/qwen_vl.py
+COPY runtime.invalid-token-failure/python/sglang/srt/managers/schedule_batch.py /sgl-workspace/sglang/python/sglang/srt/managers/schedule_batch.py
+COPY provenance/invalid-token-failure-runtime-files.json /tmp/invalid-token-failure-runtime-files.json
+RUN python3 -B -c 'import hashlib,json,pathlib; root=pathlib.Path("/sgl-workspace/sglang"); expected=json.loads(pathlib.Path("/tmp/invalid-token-failure-runtime-files.json").read_text()); actual={str(p.relative_to(root)) for p in (root/"python/sglang").rglob("*") if p.is_file() and "__pycache__" not in p.parts and p.suffix != ".pyc"}; assert actual == set(expected), (len(actual), len(expected)); bad=[n for n,h in expected.items() if hashlib.sha256((root/n).read_bytes()).hexdigest()!=h]; assert not bad, bad; files=[root/"python/sglang/srt/entrypoints/openai"/n for n in ("serving_chat.py", "serving_completions.py", "protocol.py", "serving_responses.py", "responses_compat.py")]+[root/"python/sglang/srt/function_call/qwen3_coder_detector.py", root/"python/sglang/srt/multimodal/processors/qwen_vl.py", root/"python/sglang/srt/managers/schedule_batch.py"]; [compile(f.read_bytes(), str(f), "exec") for f in files]' \
+ && rm /tmp/invalid-token-failure-runtime-files.json
+ENTRYPOINT ["python3", "-m", "sglang.launch_server"]
+CMD ["--help"]
diff --git a/Dockerfile.qwen-multimodal-alias b/Dockerfile.qwen-multimodal-alias
@@ -0,0 +1,17 @@
+# Cumulative CPU-only overlay: patches 0015, 0016, 0017, then 0018.
+# All model paths, serving arguments, and runtime settings remain external.
+FROM docker.io/kanadaj/sglang-qwen38fn-sm120-turbo@sha256:872a2bda228e39aa9c1af729b47cc28f7862e7859e448f1a8868b85a4051f404
+ARG SOURCE_REVISION
+LABEL org.opencontainers.image.source="https://github.com/kanadaj/sglang" \
+      org.opencontainers.image.revision="${SOURCE_REVISION}"
+COPY runtime/python/sglang/srt/entrypoints/openai/serving_chat.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_chat.py
+COPY runtime/python/sglang/srt/entrypoints/openai/protocol.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/protocol.py
+COPY runtime/python/sglang/srt/entrypoints/openai/serving_responses.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/serving_responses.py
+COPY runtime/python/sglang/srt/entrypoints/openai/responses_compat.py /sgl-workspace/sglang/python/sglang/srt/entrypoints/openai/responses_compat.py
+COPY runtime/python/sglang/srt/function_call/qwen3_coder_detector.py /sgl-workspace/sglang/python/sglang/srt/function_call/qwen3_coder_detector.py
+COPY runtime/python/sglang/srt/multimodal/processors/qwen_vl.py /sgl-workspace/sglang/python/sglang/srt/multimodal/processors/qwen_vl.py
+COPY provenance/qwen-multimodal-alias-runtime-files.json /tmp/qwen-multimodal-alias-runtime-files.json
+RUN python3 -B -c 'import hashlib,json,pathlib; root=pathlib.Path("/sgl-workspace/sglang"); expected=json.loads(pathlib.Path("/tmp/qwen-multimodal-alias-runtime-files.json").read_text()); actual={str(p.relative_to(root)) for p in (root/"python/sglang").rglob("*") if p.is_file() and "__pycache__" not in p.parts and p.suffix != ".pyc"}; assert actual == set(expected), (len(actual), len(expected)); bad=[n for n,h in expected.items() if hashlib.sha256((root/n).read_bytes()).hexdigest()!=h]; assert not bad, bad; files=[root/"python/sglang/srt/entrypoints/openai"/n for n in ("serving_chat.py", "protocol.py", "serving_responses.py", "responses_compat.py")]+[root/"python/sglang/srt/function_call/qwen3_coder_detector.py", root/"python/sglang/srt/multimodal/processors/qwen_vl.py"]; [compile(f.read_bytes(), str(f), "exec") for f in files]' \
+ && rm /tmp/qwen-multimodal-alias-runtime-files.json
+ENTRYPOINT ["python3", "-m", "sglang.launch_server"]
+CMD ["--help"]
diff --git a/Dockerfile.responses-compat b/Dockerfile.responses-compat
@@ -1,4 +1,4 @@
-# API-only compatibility overlay. All serving arguments remain external.
+# API-only compatibility and phase/order overlay. All serving arguments remain external.
 # Immutable rollback/base; no private snapshot, checkpoint or docker commit.
 FROM docker.io/kanadaj/sglang-qwen38fn-sm120-turbo@sha256:872a2bda228e39aa9c1af729b47cc28f7862e7859e448f1a8868b85a4051f404
 ARG SOURCE_REVISION

diff --git a/README.md b/README.md
@@ -1,8 +1,23 @@
 # Qwen TP2 packed-PLE vision on SM120
 
-**Unpublished CPU candidate:** [Responses namespace/custom compatibility](docs/responses-compat.md)
-adds a separately attested boundary backport after the effort-alias profile.
-Historical production profiles below are unchanged; no deployment is implied.
+**Published cumulative compatibility runtime:**
+[`production-cumulative-compat-20260914-v3`](docs/production-cumulative-compat-20260914.md)
+contains the ordered `0015` → `0016` → `0017` → `0018` → `0019` stack at
+`docker.io/kanadaj/sglang-qwen38fn-sm120-turbo@sha256:f2859d1ccf824a5295088cf578eba89b0f3eeefff6ae7679c3f5d64af0689458`.
+It includes the Qwen effort aliases, [Responses namespace/custom and phase/order
+compatibility](docs/responses-compat.md), [Flash-Next multimodal processor
+aliases](docs/qwen-multimodal-alias.md), and [invalid generated-token failure
+propagation](docs/invalid-token-failure.md). The image was rebuilt from clean
+`main`, verified against all 4,392 source hashes, and anonymously pulled by tag
+and digest. Publication does not imply production deployment.
+
+**Opt-in HiCache profile:** [Qwen RAM/file state transfer](docs/hicache-wip.md)
+adds companion-state handling and upstream QSA, router and restore-order fixes
+after the cumulative runtime. See the configuration-specific live results and
+remaining limits before enabling it; default images and patch series are unchanged.
+
+Historical production profiles below remain available as rollback and audit
+records.
 
 Publishable source and deployment package for the locally accepted Qwen3.8
 Flash-Next LIL NVFP4 stack. **No model weights, container archives, credentials,