Document and package DCP partial-batch runtime fix

Dockerfile.runtime

@@ -7,3 +7,8 @@ COPY runtime/glm52_vision \
   /opt/venv/lib/python3.12/site-packages/glm52_vision
 COPY runtime/glm52_vision_vllm-0.1.0.dist-info \
   /opt/venv/lib/python3.12/site-packages/glm52_vision_vllm-0.1.0.dist-info
+
+# Preserve the fast DCP project-before-merge path for partial prefill batches.
+COPY patch_dcp_workspace_stride.py /tmp/patch_dcp_workspace_stride.py
+RUN python3 /tmp/patch_dcp_workspace_stride.py \
+    && rm /tmp/patch_dcp_workspace_stride.py

README.md

@@ -62,15 +62,16 @@ as multimodal.
 | Recipe ID | `glm-5.2-exl3` |
 | Served model | `GLM-5.2-EXL3` |
 | Checkpoint | `GLM-5.2-EXL3-TR3-3.0bpw` |
-| Runtime image | `local/glm52-exl3-vision:v21` |
+| Runtime image | `local/glm52-exl3-vision:v22-dcpstride` |
 | Parallelism | TP4, DCP4, `ag_rs` |
 | Context | 400,000 tokens |
-| KV cache | FP8 MLA, 886,756 tokens, 2.22x concurrency at 400K |
+| KV cache | FP8 MLA, 878,822 tokens, 2.20x concurrency at 400K |
 | Scheduler | async, 2,048 max batched tokens, 8 max sequences |
 | Speculation | MTP-3, greedy, EXL3 TR3 3.0 bpw layer 78 |
 | Collectives | NCCL; B12X PCIe all-reduce and DCP A2A disabled |
 | Graph mode | `FULL_AND_PIECEWISE` CUDA graphs |
 | API features | `glm45` reasoning, `glm47` tools |
+| Runtime fix | pitched DCP workspace validation for partial prefill batches |
 
 ### `EXL3-VISION` — multimodal profile
 
@@ -94,6 +95,12 @@ as multimodal.
 Both profiles used the publisher-tested physical rank order `3,1,2,0`.
 Every GPU remained capped at 300 W.
 
+The text runtime adds a one-line layout validation fix for partial DCP prefill
+batches. The projection input was already compacted before `torch.bmm`; the
+validator now accepts the actual pitched stride of the 2,048-token workspace.
+This keeps DCP4, project-before-merge, and CUDA graphs enabled under concurrent
+mixed-length requests.
+
 ## Measured speed: text-only
 
 Hardware: 4 x RTX PRO 6000 96 GB, power limit 300 W/GPU. The canonical decode

docker-compose.yml

@@ -3,7 +3,7 @@
 # the merged Glm5v checkpoint and the packaged MoonViT/PatchMerger vLLM plugin.
 services:
   glm52:
-    image: ${IMAGE:-local/glm52-exl3-vision:v21}
+    image: ${IMAGE:-local/glm52-exl3-vision:v22-dcpstride}
     container_name: ${CONTAINER_NAME:-glm52-tr3-vision}
     ports:
       - "${BIND_ADDRESS:-127.0.0.1}:${PORT:-8000}:8000"

patch_dcp_workspace_stride.py

@@ -0,0 +1,22 @@
+from pathlib import Path
+
+path = Path(
+    "/opt/venv/lib/python3.12/site-packages/vllm/v1/attention/"
+    "backends/mla/b12x_mla_sparse.py"
+)
+source = path.read_text()
+old = """        expected_attn_stride = (
+            self.kv_lora_rank,
+            (self._max_batched if self._pad_heads else num_tokens) * self.kv_lora_rank,
+            1,
+        )
+"""
+new = """        expected_attn_stride = (
+            self.kv_lora_rank,
+            self._max_batched * self.kv_lora_rank,
+            1,
+        )
+"""
+if old not in source:
+    raise SystemExit("expected DCP workspace stride source was not found")
+path.write_text(source.replace(old, new, 1))

start.sh

@@ -5,7 +5,7 @@ set -Eeuo pipefail
 readonly MODEL_REPO="${MODEL_REPO:-0xSero/GLM-5.2-TR3-Vision}"
 readonly MODEL_DIR="${MODEL_DIR:-$PWD/GLM-5.2-TR3-Vision}"
 readonly CACHE_DIR="${CACHE_DIR:-$MODEL_DIR/.runtime-cache}"
-readonly IMAGE="${IMAGE:-local/glm52-exl3-vision:v21}"
+readonly IMAGE="${IMAGE:-local/glm52-exl3-vision:v22-dcpstride}"
 
 command -v docker >/dev/null
 docker compose version >/dev/null