Port SGLang integration to v0.5.15.post1 APIs

README.md

@@ -82,12 +82,32 @@ Foundry ships engine integrations under `foundry/python/foundry/integration/`. P
 | Engine | Single GPU | DP | TP | EP |
 |---|:---:|:---:|:---:|:---:|
 | vLLM | ✅ | ✅ | 🚧 | ✅ |
-| SGLang | ✅ | ✅ | 🚧 | ✅ |
+| SGLang `0.5.15.post1` | ✅ | ✅ | ✅[^sglang-tp] | ✅ |
 | TensorRT-LLM | 🚧 | 🚧 | 🚧 | 🚧 |
 
 ✅ validated end-to-end (SAVE → LOAD → query)  ·  🚧 not yet
 
-The adapted vLLM / SGLang / TensorRT-LLM forks will be released alongside this repo at `foundry-org/vllm`, `foundry-org/sglang`, `foundry-org/TensorRT-LLM`.
+[^sglang-tp]: SGLang conventional tensor parallelism is validated only for
+    dense TP=2 × DP=2 with PP=1 and EP=1. All other conventional TP topologies
+    remain unsupported.
+
+The SGLang integration is latest-only for upstream `v0.5.15.post1`
+(`0b3bb0c`), CUDA 13.0, and PyTorch 2.11/cu130 on H100. Full decode with
+prefill graphs disabled is validated for single-GPU FlashInfer and regular
+DP=2 FlashInfer on Qwen3-1.7B (52 graphs/rank), exact dense TP=2 × DP=2
+FlashInfer on Qwen3-8B (36 graphs/rank), and DP-attention EP=2 and EP=4 with
+DeepEP low-latency, DeepGEMM, and FA3 on Qwen3-30B-A3B-FP8 (20 graphs/rank).
+All passed baseline/SAVE/fresh LOAD/query with text/token parity and no
+recapture. EP=4 uses DP-attention/EP rather than conventional TP attention,
+even though SGLang's process topology sets `tp_size=4`. Existing SGLang
+archives are incompatible and require a fresh SAVE. See
+[`recipe/sglang/README.md`](recipe/sglang/README.md).
+
+For the current Foundry-only SGLang workflow, check out upstream
+`sgl-project/sglang` at the tagged release and apply the documented four-file
+patch. `foundry-org/sglang` should carry that same patch when its release branch
+is published; the current instructions do not assume that branch already
+exists. Adapted engine forks will be released alongside this repository.
 
 ### Performance
 

ROADMAP.md

@@ -30,8 +30,8 @@
     - [x] Quantized MoE with DeepGemm
 - [x] Drop-in integration layer for CUDA graph persistence in vLLM
 - [x] Sync with latest SGLang release
-    - [ ] EP on SGLang
-    - [ ] TP on SGLang
+    - [x] EP on SGLang (DP-attention + DeepEP, validated at EP=2 and EP=4)
+    - [x] Exact dense TP=2 × DP=2 on SGLang (all other TP topologies unsupported)
 
 ## Stage 5: Disaggregated and Large-Scale Serving