fix(recipes): glm-5.2: move SGLang B200/H200 recipes to stock Dynamo 1.5.1 (SGLang 0.5.18, Mooncake PD, H200 decode KV fix)

docs/fern/pages/recipes/_catalog/recipes/glm-5-2.yaml

@@ -11,11 +11,10 @@ model:
   precision: NVFP4 + FP8 KV (B200) / FP8 + FP8 KV (H200)
 status: validated
 artifacts:
-  recipe_specific_images:
-  - nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.3.0-glm-5.2-dev.1
   recipe_specific_image_periods:
   - image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.3.0-glm-5.2-dev.1
     source_revision: 9ab57d7ecefdd2a2af2e2a2c889724a157457cd6
+    effective_to: "2026-10-04"
     source_kind: github-release
     release_tag: v1.3.0-glm-5.2-dev.1
     release_state: prerelease
@@ -24,9 +23,10 @@ targets:
   recommended: false
   hardware:
     gpu: B200
-    count: 4
+    count: 16
   runtime:
     framework: sglang
+    image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.5.1
   topology: aggregated
   techniques:
   - nvfp4
@@ -44,14 +44,15 @@ targets:
     asset: recipes/glm-5.2/perf/perf.yaml
   expected_performance:
     available: true
-    summary: 176.42 output tok/s/GPU, 57.49 user tok/s P50, TTFT P50 355.6 ms at concurrency 64 (15% agentic trace subset)
+    summary: 190.05 output tok/s/GPU, 61.92 user tok/s P50, TTFT P50 228.7 ms at concurrency 64 (4 workers, 15% agentic trace subset)
 - id: sglang-disagg-b200
   recommended: false
   hardware:
     gpu: B200
-    count: 12
+    count: 20
   runtime:
     framework: sglang
+    image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.5.1
   topology: disaggregated
   techniques:
   - disaggregated-prefill-decode
@@ -60,7 +61,7 @@ targets:
   - eagle-mtp-spec-dec
   - kv-aware-routing
   - hicache-cpu-offload
-  - nixl-ucx
+  - mooncake
   workload:
     type: agentic-trace-replay
     traffic: Mooncake agentic trace, 64K median ISL / 400 median OSL, 90% KV cache hit, concurrency 128
@@ -71,14 +72,15 @@ targets:
     asset: recipes/glm-5.2/perf/perf.yaml
   expected_performance:
     available: true
-    summary: 320.91 output tok/s/GPU, 65.11 user tok/s P50, TTFT P50 1938.1 ms at concurrency 128 (3P1D, 15% agentic trace subset)
+    summary: 323.82 output tok/s/GPU, 63.13 user tok/s P50, TTFT P50 1280.1 ms at concurrency 128 (3P1D, 15% agentic trace subset)
 - id: sglang-agg-h200
   recommended: false
   hardware:
     gpu: H200
-    count: 8
+    count: 24
   runtime:
     framework: sglang
+    image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.5.1
   topology: aggregated
   techniques:
   - fp8
@@ -95,22 +97,23 @@ targets:
     asset: recipes/glm-5.2/perf/perf.yaml
   expected_performance:
     available: true
-    summary: 54.55 output tok/s/GPU, 52.37 user tok/s P50, TTFT P50 1790 ms at concurrency 32 (3 workers, 15% agentic trace subset)
+    summary: 60.87 output tok/s/GPU, 57.33 user tok/s P50, TTFT P50 1158.2 ms at concurrency 32 (3 workers, 15% agentic trace subset)
 - id: sglang-disagg-h200
   recommended: false
   hardware:
     gpu: H200
     count: 16
   runtime:
     framework: sglang
+    image: nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.5.1
   topology: disaggregated
   techniques:
   - disaggregated-prefill-decode
   - fp8
   - fp8-kv-cache
   - eagle-mtp-spec-dec
   - kv-aware-routing
-  - nixl-ucx
+  - mooncake
   workload:
     type: agentic-trace-replay
     traffic: Mooncake agentic trace, 64K median ISL / 400 median OSL, 90% KV cache hit, concurrency 24
@@ -121,7 +124,7 @@ targets:
     asset: recipes/glm-5.2/perf/perf.yaml
   expected_performance:
     available: true
-    summary: 68.86 output tok/s/GPU, 53.88 user tok/s P50, TTFT P50 1874 ms at concurrency 24 (1P1D, 15% agentic trace subset)
+    summary: 84.33 output tok/s/GPU, 61.46 user tok/s P50, TTFT P50 1309.6 ms at concurrency 24 (1P1D, 15% agentic trace subset)
 related_benchmarks: []
 maintainer: milesial
 gaps:

docs/fern/pages/recipes/model-recipes/glm-5-2.mdx

@@ -30,23 +30,23 @@ Each target below is a Dynamo + SGLang deployment of Z.AI's GLM-5.3/5.2 tuned fo
 <div className="dynamo-target-picker-summary" data-sku="b200" data-variant="agg">
 <span><b>Checkpoint</b> RadixArk/GLM-5.3-NVFP4 / nvidia/GLM-5.2-NVFP4</span>
 <span><b>Precision</b> NVFP4 + FP8 KV cache</span>
-<span><b>GPUs</b> 4x B200, one worker</span>
+<span><b>GPUs</b> 16x B200 (4 workers x 4)</span>
 <span><b>Parallelism</b> DTP4</span>
 <span><b>Routing</b> KV-aware</span>
 <span><b>Context</b> Up to 500K, HiCache CPU offload</span>
 </div>
 <div className="dynamo-target-picker-summary" data-sku="b200" data-variant="disagg">
 <span><b>Checkpoint</b> RadixArk/GLM-5.3-NVFP4 / nvidia/GLM-5.2-NVFP4</span>
 <span><b>Precision</b> NVFP4 + FP8 KV cache</span>
-<span><b>GPUs</b> 4x B200 prefill + 8x B200 decode</span>
+<span><b>GPUs</b> 20x B200 (3 prefill x 4 + 1 decode x 8)</span>
 <span><b>Parallelism</b> DEP4 prefill / DTP8 decode</span>
-<span><b>Routing</b> KV-aware, NIXL/UCX over IB</span>
+<span><b>Routing</b> KV-aware, Mooncake over IB</span>
 <span><b>Context</b> Up to 500K, HiCache CPU offload</span>
 </div>
 <div className="dynamo-target-picker-summary" data-sku="h200" data-variant="agg">
 <span><b>Checkpoint</b> zai-org/GLM-5.3 / zai-org/GLM-5.2-FP8</span>
 <span><b>Precision</b> FP8 + FP8 KV cache</span>
-<span><b>GPUs</b> 8x H200, one worker</span>
+<span><b>GPUs</b> 24x H200 (3 workers x 8)</span>
 <span><b>Parallelism</b> TP8/EP8</span>
 <span><b>Routing</b> KV-aware</span>
 <span><b>Context</b> Up to 250K</span>
@@ -56,7 +56,7 @@ Each target below is a Dynamo + SGLang deployment of Z.AI's GLM-5.3/5.2 tuned fo
 <span><b>Precision</b> FP8 + FP8 KV cache</span>
 <span><b>GPUs</b> 8x H200 prefill + 8x H200 decode</span>
 <span><b>Parallelism</b> TP8/EP8 prefill / TP8/DP8/EP1 decode</span>
-<span><b>Routing</b> KV-aware, NIXL/UCX over IB</span>
+<span><b>Routing</b> KV-aware, Mooncake over IB</span>
 <span><b>Context</b> Up to 250K</span>
 </div>
 </div>
@@ -65,14 +65,14 @@ Each target below is a Dynamo + SGLang deployment of Z.AI's GLM-5.3/5.2 tuned fo
 
 <div data-sku="b200">
 
-- A Kubernetes cluster with the Dynamo platform installed and B200 GPUs available — 4x for aggregated, 12x for disaggregated. See the [Kubernetes Deployment Guide](../../kubernetes/getting-started/quickstart.mdx).
+- A Kubernetes cluster with the Dynamo platform installed and B200 GPUs available — 16x for aggregated, 20x for disaggregated. See the [Kubernetes Deployment Guide](../../kubernetes/getting-started/quickstart.mdx).
 - A Hugging Face token with access to `RadixArk/GLM-5.3-NVFP4` / `nvidia/GLM-5.2-NVFP4`.
 
 </div>
 
 <div data-sku="h200">
 
-- A Kubernetes cluster with the Dynamo platform installed and H200 GPUs available — 8x for aggregated, 16x for disaggregated. See the [Kubernetes Deployment Guide](../../kubernetes/getting-started/quickstart.mdx).
+- A Kubernetes cluster with the Dynamo platform installed and H200 GPUs available — 24x for aggregated, 16x for disaggregated. See the [Kubernetes Deployment Guide](../../kubernetes/getting-started/quickstart.mdx).
 - A Hugging Face token with access to `zai-org/GLM-5.3` / `zai-org/GLM-5.2-FP8`.
 
 </div>
@@ -102,7 +102,7 @@ kubectl apply -f recipes/glm-5.2/model-cache/model-download.yaml -n ${NAMESPACE}
 kubectl wait --for=condition=Complete job/model-download -n ${NAMESPACE} --timeout=7200s
 ```
 
-When serving GLM-5.3, update every `model-path` in the target DGD to `RadixArk/GLM-5.3-NVFP4` for B200 or `zai-org/GLM-5.3` for H200, and update every `served-model-name` to `zai-org/GLM-5.3`.
+When serving GLM-5.2, update every `model-path` in the target DGD to `nvidia/GLM-5.2-NVFP4` for B200 or `zai-org/GLM-5.2-FP8` for H200, and update every `served-model-name` to `zai-org/GLM-5.2`.
 
 Then deploy the target DGD:
 
@@ -195,7 +195,7 @@ curl http://localhost:8000/v1/chat/completions \
 
 A single AIPerf trace-replay Job — `perf/perf.yaml` — covers all four DGDs. It replays a Mooncake-format agentic trace (64K ISL / 400 OSL, 90% KV cache hit rate) at one concurrency value and writes artifacts to the shared `model-cache` PVC. The benchmark pod is co-located with a DGD frontend through `podAffinity`.
 
-Edit the `env` block in `perf/perf.yaml` to target your deployed DGD — set `ENDPOINT` to the matching frontend service, `SYNTHESIS_MAX_ISL` to its context limit, and `CONCURRENCY` to the value for that target. The `CONCURRENCY` values below reproduce the [Expected Performance (run on GLM-5.2)](#expected-performance-run-on-glm-52) numbers:
+Edit the `env` block in `perf/perf.yaml` to target your deployed DGD — set `ENDPOINT` to the matching frontend service, `SYNTHESIS_MAX_ISL` to its context limit, and `CONCURRENCY` to the value for that target. The `CONCURRENCY` values below reproduce the [Expected Performance (run on GLM-5.3)](#expected-performance-run-on-glm-53) numbers. The rows also require `SGLANG_SIMULATE_ACC_LEN=2.69`, `SGLANG_SIMULATE_ACC_METHOD=match-expected`, and `SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token` uncommented on the aggregated workers or the disaggregated decode workers (see [perf/README.md](https://github.com/ai-dynamo/dynamo/blob/main/recipes/glm-5.2/perf/README.md)); keep them commented for accuracy evaluation and production:
 
 | Target | `ENDPOINT` | `SYNTHESIS_MAX_ISL` | `CONCURRENCY` |
 | --- | --- | --- | --- |
@@ -213,13 +213,19 @@ kubectl wait --for=condition=Complete job/glm52-bench -n ${NAMESPACE} --timeout=
 
 For trace staging, concurrency sweeps, and fetching artifacts, see the [benchmark README](https://github.com/ai-dynamo/dynamo/blob/main/recipes/glm-5.2/perf/README.md).
 
-## Expected Performance (run on GLM-5.2)
+## Expected Performance (run on GLM-5.3)
 
 Measured on the 15% agentic trace subset. System throughput is per-GPU output tokens per second; user throughput is the P50 per-request output rate.
 
 | | Concurrency | System output tok/s/GPU | User output tok/s (P50) | TTFT P50 (ms) |
 |---|---|---|---|---|
-| <span data-sku="b200" data-variant="agg">**B200 aggregated** (4 workers)</span><span data-sku="b200" data-variant="disagg">**B200 disaggregated** (3P1D)</span><span data-sku="h200" data-variant="agg">**H200 aggregated** (3 workers)</span><span data-sku="h200" data-variant="disagg">**H200 disaggregated** (1P1D)</span> | <span data-sku="b200" data-variant="agg">64</span><span data-sku="b200" data-variant="disagg">128</span><span data-sku="h200" data-variant="agg">32</span><span data-sku="h200" data-variant="disagg">24</span> | <span data-sku="b200" data-variant="agg">176.420</span><span data-sku="b200" data-variant="disagg">320.907</span><span data-sku="h200" data-variant="agg">54.550</span><span data-sku="h200" data-variant="disagg">68.860</span> | <span data-sku="b200" data-variant="agg">57.493</span><span data-sku="b200" data-variant="disagg">65.105</span><span data-sku="h200" data-variant="agg">52.370</span><span data-sku="h200" data-variant="disagg">53.880</span> | <span data-sku="b200" data-variant="agg">355.555</span><span data-sku="b200" data-variant="disagg">1938.059</span><span data-sku="h200" data-variant="agg">1790.000</span><span data-sku="h200" data-variant="disagg">1874.000</span> |
+| <span data-sku="b200" data-variant="agg">**B200 aggregated** (4 workers)</span><span data-sku="b200" data-variant="disagg">**B200 disaggregated** (3P1D)</span><span data-sku="h200" data-variant="agg">**H200 aggregated** (3 workers)</span><span data-sku="h200" data-variant="disagg">**H200 disaggregated** (1P1D)</span> | <span data-sku="b200" data-variant="agg">64</span><span data-sku="b200" data-variant="disagg">128</span><span data-sku="h200" data-variant="agg">32</span><span data-sku="h200" data-variant="disagg">24</span> | <span data-sku="b200" data-variant="agg">190.048</span><span data-sku="b200" data-variant="disagg">323.821</span><span data-sku="h200" data-variant="agg">60.866</span><span data-sku="h200" data-variant="disagg">84.335</span> | <span data-sku="b200" data-variant="agg">61.923</span><span data-sku="b200" data-variant="disagg">63.133</span><span data-sku="h200" data-variant="agg">57.330</span><span data-sku="h200" data-variant="disagg">61.460</span> | <span data-sku="b200" data-variant="agg">228.700</span><span data-sku="b200" data-variant="disagg">1280.100</span><span data-sku="h200" data-variant="agg">1158.200</span><span data-sku="h200" data-variant="disagg">1309.600</span> |
+
+<div data-sku="h200">
+
+3,535 of 3,541 trace requests completed; 6 requests exceed the 250K context limit.
+
+</div>
 
 ## Compare All Targets
 
@@ -229,21 +235,23 @@ All four targets serve GLM-5.3/5.2 on SGLang with KV-aware routing and EAGLE-sty
 |---|---|---|---|---|
 | **Checkpoint** | RadixArk/GLM-5.3-NVFP4 / nvidia/GLM-5.2-NVFP4 | RadixArk/GLM-5.3-NVFP4 / nvidia/GLM-5.2-NVFP4 | zai-org/GLM-5.3 / zai-org/GLM-5.2-FP8 | zai-org/GLM-5.3 / zai-org/GLM-5.2-FP8 |
 | **Precision** | NVFP4 + FP8 KV | NVFP4 + FP8 KV | FP8 + FP8 KV | FP8 + FP8 KV |
-| **GPUs** | 4x B200 | 4x B200 prefill + 8x B200 decode | 8x H200 | 8x H200 prefill + 8x H200 decode |
+| **GPUs** | 16x B200 (4 workers x 4) | 20x B200 (3 prefill x 4 + 1 decode x 8) | 24x H200 (3 workers x 8) | 16x H200 (8 prefill + 8 decode) |
 | **Parallelism** | DTP4 | DEP4 / DTP8 | TP8/EP8 | TP8/EP8 prefill / TP8/DP8/EP1 decode |
 | **KV offload** | HiCache CPU | HiCache CPU | None | None |
 | **Max context** | 500K | 500K | 250K | 250K |
+| **Image** | `sglang-runtime:1.5.1` | `sglang-runtime:1.5.1` | `sglang-runtime:1.5.1` | `sglang-runtime:1.5.1` |
 
 ## Notes
 
 - Speculative decoding uses EAGLE-style MTP with draft length 3, measured at a SpeedBench acceptance length of 2.69.
 - All four targets use KV-aware routing at the frontend. The B200 targets add HiCache CPU offload; the agentic traces are shaped to show the value of KV-aware routing and offloading.
-- The disaggregated targets transfer KV over NIXL/UCX on InfiniBand.
+- The disaggregated targets transfer KV over Mooncake on InfiniBand.
+- The H200 disaggregated decode worker uses `mem-fraction-static: 0.95`; at 0.88 the per-rank KV pool is smaller than one 250K-token request.
 
 ## Limitations
 
 - B200 targets support up to 500K context; the full 1M context length is not supported out of the box. H200 targets support up to 250K context.
-- Structured decoding requires reasoning to be disabled (`"chat_template_kwargs": {"enable_thinking": false}`) so the output lands in the `content` field instead of `reasoning_content`.
+- Structured decoding works with reasoning enabled: the generated JSON is populated in the `content` field and the chain-of-thought in `reasoning_content`. This requires both `--dyn-reasoning-parser glm45` (frontend) and `--reasoning-parser glm45` (engine), which the recipes set.
 - `n>1` requests are not supported with the disaggregated targets.
 
 ## Source

docs/tests/test_recipe_catalog_artifacts.py

@@ -25,7 +25,7 @@
         ),
         (
             "glm-5-2",
-            ("nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.3.0-glm-5.2-dev.1",),
+            (),
         ),
         (
             "inkling",
@@ -70,7 +70,10 @@ def test_recipe_specific_images_are_catalog_owned(
     expected_images: tuple[str, ...],
 ) -> None:
     document = yaml.safe_load((CATALOG / "recipes" / f"{recipe_id}.yaml").read_text())
-    assert tuple(document["artifacts"]["recipe_specific_images"]) == expected_images
+    assert (
+        tuple(document["artifacts"].get("recipe_specific_images", ()))
+        == expected_images
+    )
 
 
 @pytest.mark.parametrize(
@@ -94,6 +97,7 @@ def test_recipe_specific_images_are_catalog_owned(
                 {
                     "image": "nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.3.0-glm-5.2-dev.1",
                     "source_revision": "9ab57d7ecefdd2a2af2e2a2c889724a157457cd6",
+                    "effective_to": "2026-10-04",
                     "source_kind": "github-release",
                     "release_tag": "v1.3.0-glm-5.2-dev.1",
                     "release_state": "prerelease",

recipes/glm-5.2/README.md

@@ -22,7 +22,9 @@ Dynamo + SGLang deployment profiles for the B200 and H200 agentic workload:
 | **Speculative decoding** | EAGLE-style MTP (DL=3, SpeedBench AL=2.69) | EAGLE-style MTP (DL=3, SpeedBench AL=2.69) | EAGLE-style MTP (DL=3, SpeedBench AL=2.69) | EAGLE-style MTP (DL=3, SpeedBench AL=2.69) |
 | **Context length**       | 500,000                                    | 500,000                                    | 250,000                                    | 250,000                                    |
 | **KV cache offloading**  | HiCache CPU                                | HiCache CPU                                | None                                       | None                                       |
-| **KV transfer**          | N/A                                        | NIXL/UCX over IB                           | N/A                                        | NIXL/UCX over IB                           |
+| **KV transfer**          | N/A                                        | Mooncake over IB                           | N/A                                        | Mooncake over IB                           |
+
+All four variants run on `nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.5.1` (Dynamo 1.5.1, SGLang 0.5.18).
 
 
 ## Supported features
@@ -72,9 +74,9 @@ kubectl wait --for=condition=Complete job/model-download -n ${NAMESPACE} --timeo
 
 ### 4. Deploy the DGD
 
-When serving GLM-5.3, update every `model-path` in the target DGD to
-`RadixArk/GLM-5.3-NVFP4` for B200 or `zai-org/GLM-5.3` for H200, and update every
-`served-model-name` to `zai-org/GLM-5.3`.
+When serving GLM-5.2, update every `model-path` in the target DGD to
+`nvidia/GLM-5.2-NVFP4` for B200 or `zai-org/GLM-5.2-FP8` for H200, and update every
+`served-model-name` to `zai-org/GLM-5.2`.
 
 Deploy the target DGD:
 
@@ -100,15 +102,17 @@ See [perf/README.md](perf/README.md) for the full benchmark workflow — trace s
 
 Modified Mooncake traces are provided to showcase the value of KV-aware routing and CPU offloading, see [perf/README.md](perf/README.md) for details.
 
-## Performance results (run on GLM-5.2)
+## Performance results (run on GLM-5.3)
+
 
+| Workload             | Framework | Recipe                 | SKU  | GPUs | Concurrency | System output tok/s/GPU | User output tok/s (P50) | TTFT P50 (ms) |
+| -------------------- | --------- | ---------------------- | ---- | ---- | ----------- | ----------------------- | ----------------------- | ------------- |
+| Agentic (15% subset) | SGLang    | Aggregated (4 workers) | B200 | 16   | 64          | 190.048                 | 61.923                  | 228.700       |
+| Agentic (15% subset) | SGLang    | Disaggregated (3P1D)   | B200 | 20   | 128         | 323.821                 | 63.133                  | 1280.100      |
+| Agentic (15% subset) | SGLang    | Aggregated (3 workers) | H200 | 24   | 32          | 60.866                  | 57.330                  | 1158.200      |
+| Agentic (15% subset) | SGLang    | Disaggregated (1P1D)   | H200 | 16   | 24          | 84.335                  | 61.460                  | 1309.600      |
 
-| Workload             | Recipe                 | SKU  | Concurrency | System output tok/s/gpu | User output tok/s (P50) | TTFT P50 (ms) |
-| -------------------- | ---------------------- | ---- | ----------- | ----------------------- | ----------------------- | ------------- |
-| Agentic (15% subset) | Aggregated (4 workers) | B200 | 64          | 176.420                 | 57.493                  | 355.555       |
-| Agentic (15% subset) | Disaggregated (3P1D)   | B200 | 128         | 320.907                 | 65.105                  | 1938.059      |
-| Agentic (15% subset) | Aggregated (3 workers) | H200 | 32          | 54.550                  | 52.370                  | 1790.000      |
-| Agentic (15% subset) | Disaggregated (1P1D)   | H200 | 24          | 68.860                  | 53.880                  | 1874.000      |
+The rows require `SGLANG_SIMULATE_ACC_LEN=2.69`, `SGLANG_SIMULATE_ACC_METHOD=match-expected`, and `SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token` uncommented on the aggregated workers or the disaggregated decode workers (see [perf/README.md](perf/README.md)); keep them commented for accuracy evaluation and production. H200 rows: 3,535 of 3,541 trace requests completed; 6 requests exceed the 250K context limit.