fix(patch): repair spec-capture-ascend-mount.patch hunk headers

docs/basic_usage/Ascend/ascend_npu.md

@@ -23,6 +23,42 @@ python -m pip install -e . --no-deps
 torch/sglang. If a later step reports a missing lightweight dependency, install
 just that package, also with `--no-deps`.
 
+### Install Mooncake (Ascend)
+
+Online runs move features through Mooncake; the launcher checks for the
+`mooncake_master` binary and the `mooncake` Python package at startup. The
+fastest path is the official NPU wheel:
+
+```bash
+pip install mooncake-transfer-engine-npu
+```
+
+If the wheel does not cover your platform, build from source instead (see the
+[vLLM-Ascend Mooncake guide](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/features/pd_colocated_mooncake_multi_instance.html)
+for the full walkthrough):
+
+```bash
+git clone -b v0.3.9 --depth 1 https://github.com/kvcache-ai/Mooncake.git
+cd Mooncake
+git submodule update --init --recursive
+apt-get install -y mpich libmpich-dev
+bash dependencies.sh -y
+mkdir build && cd build
+cmake .. -DUSE_ASCEND_DIRECT=ON   # mandatory on Ascend
+make -j && make install
+```
+
+Either way, verify both pieces before continuing:
+
+```bash
+mooncake_master --help >/dev/null && echo mooncake_master ok
+python -c "from mooncake.store import MooncakeDistributedStore; print('mooncake.store ok')"
+```
+
+If the capture server later dies with "no `allocate_and_mount_segment`", the
+Mooncake build is too old for the companion patch below — upgrade or rebuild
+it first.
+
 ### Apply the SGLang capture patches (online runs only)
 
 Online capture needs two patches on top of the installed SGLang, applied **in
@@ -35,7 +71,7 @@ bash scripts/apply_sglang_spec_capture_patch.sh
 # Ascend companion patch: skip the wildcard segment mount that Ascend
 # Mooncake rejects, and mount the feature segment with location="cpu"
 SGLANG_DIR=$(python -c "import sglang, os; print(os.path.dirname(os.path.dirname(sglang.__file__)))")
-cd "$SGLANG_DIR" && git apply /path/to/SpecForge/patches/sglang/v0.5.14/spec-capture-ascend-mount.patch
+cd "$SGLANG_DIR" && git apply -p2 /path/to/SpecForge/patches/sglang/v0.5.14/spec-capture-ascend-mount.patch
 ```
 
 Skip both for offline training, which reads features from disk. The companion

patches/sglang/v0.5.14/spec-capture-ascend-mount.patch

@@ -1,8 +1,8 @@
 diff --git a/python/sglang/srt/spec_capture_sink.py b/python/sglang/srt/spec_capture_sink.py
-index 4955e11..902f0b1 100644
+index d54b920..f831b03 100644
 --- a/python/sglang/srt/spec_capture_sink.py
 +++ b/python/sglang/srt/spec_capture_sink.py
-@@ -81,18 +81,31 @@ class SpecCaptureSink:
+@@ -87,18 +87,28 @@ class SpecCaptureSink:
              from mooncake.store import MooncakeDistributedStore, ReplicateConfig
 
              store = MooncakeDistributedStore()
@@ -38,7 +38,7 @@ index 4955e11..902f0b1 100644
                  rdma_devices=os.environ.get("MOONCAKE_RDMA_DEVICES", ""),
                  master_server_addr=os.environ.get(
                      "MOONCAKE_MASTER_SERVER_ADDR", "localhost:50051"
-@@ -100,12 +113,34 @@ class SpecCaptureSink:
+@@ -106,6 +116,25 @@ class SpecCaptureSink:
              )
              if rc is not None and int(rc) != 0:
                  raise RuntimeError(f"spec-capture mooncake setup failed (status {rc})")
@@ -64,13 +64,3 @@ index 4955e11..902f0b1 100644
              # Hard-pin every object: SpecForge (not Mooncake's LRU) is the
              # lifetime authority — a committed feature must never be evicted
              # before the trainer consumes it.
-             cfg = ReplicateConfig()
-             cfg.replica_num = 1
--            cfg.with_hard_pin = True
-+            # Older Mooncake builds expose no with_hard_pin field; objects
-+            # then follow the store's default pin behavior until remove().
-+            if hasattr(cfg, "with_hard_pin"):
-+                cfg.with_hard_pin = True
-             self._put_config = cfg
-             self._store = store
-             logger.info("spec-capture mooncake sink connected")