recipes/docker-images/mia-tp4/scripts/roce_ring-sparkring-f16b5f4.patch

diff -ruN a/roce_ring/__init__.py b/roce_ring/__init__.py
--- a/roce_ring/__init__.py
+++ b/roce_ring/__init__.py
@@ -1,5 +1,9 @@
-"""``b12x.comm.roce`` (RoCEnante): one-shot all-reduce and all-gather over RoCE for multi-node TP.
+"""``b12x.comm.roce_ring`` (RoCEnante with per-peer paths): one-shot all-reduce and all-gather over RoCE.
 
+The per-peer path variant of ``b12x.comm.roce`` for a switchless four-node ring whose opposite
+nodes are reached through hardware-forwarded paths (FujitsuPolycom/sparkring); see
+``runtime/b12x/roce_ring-provenance.json`` in the recipe for origin and local changes.
+
 Target: clusters of DGX Spark nodes joined by their ConnectX-7 200 GbE ports,
 one GPU per node.  The GB10's unified memory lets the NIC RDMA-write straight
 into pinned host memory that the GPU kernel then reads in place, so no
@@ -19,7 +23,7 @@
 from ..._lib.meta import OpMeta, Provenance, install_lazy_api
 
 META = OpMeta(
-    name="roce",
+    name="roce_ring",
     group="comm",
     api_style="stateful",
     entry_points=(
diff -ruN a/roce_ring/_allgather_cute.py b/roce_ring/_allgather_cute.py
--- a/roce_ring/_allgather_cute.py
+++ b/roce_ring/_allgather_cute.py
@@ -351,7 +351,7 @@
         1,
         1,
         current_cuda_stream(),
-        compile_spec=KernelCompileSpec.from_key("comm.roce.allgather", 2, cache_key),
+        compile_spec=KernelCompileSpec.from_key("comm.roce_ring.allgather", 2, cache_key),
     )
 
     def run(
diff -ruN a/roce_ring/_oneshot_cute.py b/roce_ring/_oneshot_cute.py
--- a/roce_ring/_oneshot_cute.py
+++ b/roce_ring/_oneshot_cute.py
@@ -429,7 +429,7 @@
         1,
         1,
         current_cuda_stream(),
-        compile_spec=KernelCompileSpec.from_key("comm.roce.oneshot", 2, cache_key),
+        compile_spec=KernelCompileSpec.from_key("comm.roce_ring.oneshot", 2, cache_key),
     )
 
     def run(
diff -ruN a/roce_ring/_path_config.py b/roce_ring/_path_config.py
--- a/roce_ring/_path_config.py
+++ b/roce_ring/_path_config.py
@@ -9,6 +9,7 @@
 PATH_COUNT = 2
 MAX_PATH_COUNT = 4
 PEER_HCA_MAP_ENV = "B12X_ROCE_PEER_HCA_MAP"
+PEER_HCA_MAPS_ENV = "B12X_ROCE_PEER_HCA_MAPS"
 OPPOSITE_PATHS_ENV = "B12X_ROCE_OPPOSITE_PATHS"
 CANONICAL_TP4_HCAS = (
     "rocep1s0f0",
@@ -67,6 +68,8 @@
     the opposite-rank entry has four distinct HCA indices. A two-HCA runtime
     defaults every non-local peer to ``0/1``; runtimes with more HCAs require
     an explicit mapping because local and reciprocal indices can differ.
+    ``B12X_ROCE_PEER_HCA_MAPS`` carries one such map per rank, ``;``-separated in rank order, so a
+    single launcher environment serves every rank; ``B12X_ROCE_PEER_HCA_MAP`` wins when both are set.
     """
 
     if hca_count < PATH_COUNT:
@@ -75,6 +78,14 @@
             f"got {hca_count}"
         )
     raw = os.getenv(PEER_HCA_MAP_ENV, "").strip()
+    per_rank = os.getenv(PEER_HCA_MAPS_ENV, "").strip()
+    if not raw and per_rank:
+        maps = [item.strip() for item in per_rank.split(";")]
+        if len(maps) != world_size:
+            raise ValueError(
+                f"{PEER_HCA_MAPS_ENV} needs {world_size} ';'-separated maps, got {len(maps)}"
+            )
+        raw = maps[rank]
     if not raw:
         if hca_count == PATH_COUNT:
             return tuple(
@@ -160,6 +171,7 @@
     "OPPOSITE_PATHS_ENV",
     "PATH_COUNT",
     "PEER_HCA_MAP_ENV",
+    "PEER_HCA_MAPS_ENV",
     "opposite_path_count",
     "peer_hca_map",
     "peer_path_count",
diff -ruN a/roce_ring/_proxy.py b/roce_ring/_proxy.py
--- a/roce_ring/_proxy.py
+++ b/roce_ring/_proxy.py
@@ -70,7 +70,7 @@
     if override:
         return Path(override)
     root = os.getenv("XDG_CACHE_HOME") or os.path.join(os.path.expanduser("~"), ".cache")
-    return Path(root) / "b12x" / "roce"
+    return Path(root) / "b12x" / "roce_ring"
 
 
 def _compiler() -> str:
@@ -79,7 +79,7 @@
         if candidate and shutil.which(candidate):
             return candidate
     raise RuntimeError(
-        "b12x.comm.roce needs a C compiler (gcc/cc/clang) and libibverbs headers "
+        "b12x.comm.roce_ring needs a C compiler (gcc/cc/clang) and libibverbs headers "
         "to build its RDMA proxy"
     )
 
diff -ruN a/roce_ring/_roce_proxy.c b/roce_ring/_roce_proxy.c
--- a/roce_ring/_roce_proxy.c
+++ b/roce_ring/_roce_proxy.c
@@ -44,7 +44,7 @@
 #define ROCE_PORT 1
 #define ROCE_SEND_DEPTH 256
 #define ROCE_ABI_VERSION 4
-#define ROCE_IDLE_SPINS 200000
+#define ROCE_IDLE_SPINS 20000000  /* dsv41: match the recipe's SG17 proxy; 200000 let the proxy nap mid-step */
 #define ROCE_DEFAULT_TWO_WAVE_THRESHOLD_BYTES 131072u
 #define ROCE_WAVE_MODE_TWO 0u
 #define ROCE_WAVE_MODE_MIXED_TWO 1u
diff -ruN a/roce_ring/api.py b/roce_ring/api.py
--- a/roce_ring/api.py
+++ b/roce_ring/api.py
@@ -1,4 +1,4 @@
-"""Public surface for comm.roce (docs in the op ``__init__``)."""
+"""Public surface for comm.roce_ring (docs in the op ``__init__``)."""
 
 from __future__ import annotations