improve(autotest): add hard_schema KVV suite and Anthropic inline system tests

autotest/configs/Qwen/Qwen2.5-7B-Instruct.yml

@@ -63,10 +63,12 @@ h:
     pytorch:
       suites:
       - toolcall
+      - hard_schema
       extra:
         tool-call-parser: qwen2d5
     turbomind:
       suites:
       - toolcall
+      - hard_schema
       extra:
         tool-call-parser: qwen2d5

autotest/configs/Qwen/Qwen3-8B-FP8.yml

@@ -17,13 +17,15 @@ h:
     pytorch:
       suites:
       - toolcall
+      - hard_schema
       - reasoning
       extra:
         tool-call-parser: qwen3
         reasoning-parser: default
     turbomind:
       suites:
       - toolcall
+      - hard_schema
       - reasoning
       extra:
         tool-call-parser: qwen3

autotest/configs/Qwen/Qwen3.5-27B.yml

@@ -75,3 +75,17 @@ h:
   quantization:
     turbomind:
     - awq
+- model_type:
+  - chat
+  - vl
+  engine_config:
+    tp: 2
+    extra:
+      prefix-cache-state-budget: 256
+      max-prefill-token-num: 64
+  backends:
+  - name: pytorch
+    communicators:
+    - nccl
+  test_coverage:
+  - prefix_cache

autotest/configs/Qwen/Qwen3.5-35B-A3B-FP8.yml

@@ -1,6 +1,6 @@
 # HF model id: Qwen/Qwen3.5-35B-A3B-FP8
 
-a100:
+h:
 - model_type:
   - chat
   - vl
@@ -10,6 +10,9 @@ a100:
   - name: turbomind
     communicators:
     - nccl
+  - name: pytorch
+    communicators:
+    - nccl
   test_coverage:
   - benchmark
   - evaluate
@@ -18,42 +21,23 @@ a100:
   - longtext_evaluate
   - mllm_evaluate
   interface:
-    turbomind:
+    pytorch:
     - suites:
       - base
       - logprob
+      - experts
       - toolcall
       - reasoning
       extra:
         logprobs-mode: raw_logprobs
+        enable-return-routed-experts: true
         tool-call-parser: qwen3coder
         reasoning-parser: default
     - suites:
       - anthropic
       extra:
         logprobs-mode: raw_logprobs
-    - suites:
-      - abort
-      extra:
-        enable-abort-handling: true
-h:
-- model_type:
-  - chat
-  - vl
-  engine_config:
-    tp: 1
-  backends:
-  - name: turbomind
-    communicators:
-    - nccl
-  test_coverage:
-  - benchmark
-  - evaluate
-  - func
-  - longtext_benchmark
-  - longtext_evaluate
-  - mllm_evaluate
-  interface:
+        enable-return-routed-experts: true
     turbomind:
     - suites:
       - base
@@ -68,29 +52,20 @@ h:
       - anthropic
       extra:
         logprobs-mode: raw_logprobs
-    - suites:
-      - abort
-      extra:
-        enable-abort-handling: true
 - model_type:
   - chat
   - vl
   engine_config:
-    tp: 2
+    tp: 1
+    extra:
+      prefix-cache-state-budget: 256
+      max-prefill-token-num: 64
   backends:
   - name: pytorch
     communicators:
     - nccl
   test_coverage:
-  - func
-  interface:
-    pytorch:
-    - suites:
-      - abort
-      extra:
-        enable-abort-handling: true
-    - suites:
-      - sleep
+  - prefix_cache
 - model_type:
   - chat
   - vl
@@ -111,11 +86,3 @@ h:
   - longtext_benchmark
   - longtext_evaluate
   - mllm_evaluate
-  interface:
-    pytorch:
-    - suites:
-      - abort
-      extra:
-        enable-abort-handling: true
-    - suites:
-      - sleep

autotest/configs/Qwen/Qwen3.5-35B-A3B.yml

@@ -23,6 +23,7 @@ a100:
       - logprob
       - experts
       - toolcall
+      - hard_schema
       - reasoning
       extra:
         logprobs-mode: raw_logprobs
@@ -45,6 +46,7 @@ a100:
       - base
       - logprob
       - toolcall
+      - hard_schema
       - reasoning
       extra:
         logprobs-mode: raw_logprobs
@@ -58,6 +60,9 @@ a100:
       - abort
       extra:
         enable-abort-handling: true
+  gen_config:
+    chat-template-kwargs:
+      enable_thinking: true
   quantization:
     turbomind:
     - kvint8
@@ -116,6 +121,7 @@ h:
       - logprob
       - experts
       - toolcall
+      - hard_schema
       - reasoning
       extra:
         logprobs-mode: raw_logprobs
@@ -138,6 +144,7 @@ h:
       - base
       - logprob
       - toolcall
+      - hard_schema
       - reasoning
       extra:
         logprobs-mode: raw_logprobs
@@ -156,6 +163,20 @@ h:
     - awq
     pytorch:
     - fp8
+- model_type:
+  - chat
+  - vl
+  engine_config:
+    tp: 2
+    extra:
+      prefix-cache-state-budget: 256
+      max-prefill-token-num: 64
+  backends:
+  - name: pytorch
+    communicators:
+    - nccl
+  test_coverage:
+  - prefix_cache
 - model_type:
   - chat
   - vl
@@ -205,8 +226,6 @@ ascend:
   - func
   - longtext_benchmark
   - mllm_evaluate
-  - prefix_cache
-  - mllm_evaluate
 - model_type:
   - chat
   - vl

autotest/configs/Qwen/Qwen3.5-397B-A17B-FP8.yml

@@ -61,11 +61,11 @@ h:
     ep: 8
     extra:
       max-batch-size: 256
-      cache-max-entry-count: 0.7
-      prefix-cache-decode-state-interval: 1024
+      cache-max-entry-count: 0.9
       prefix-cache-state-budget: 256
+      max-prefill-token-num: 64
   backends:
-  - name: turbomind
+  - name: pytorch
     communicators:
     - nccl
   test_coverage:
@@ -87,7 +87,7 @@ h:
     communicators:
     - nccl
   test_coverage:
-  - prefix_cache
+  - prefix_cache_evaluate
 - model_type:
   - chat
   - vl

autotest/configs/Qwen/Qwen3.5-397B-A17B.yml

@@ -71,20 +71,32 @@ h:
     extra:
       max-batch-size: 256
       cache-max-entry-count: 0.7
-      prefix-cache-decode-state-interval: 1024
       prefix-cache-state-budget: 256
+      max-prefill-token-num: 64
   backends:
-  - name: turbomind
+  - name: pytorch
     communicators:
     - nccl
+  test_coverage:
+  - prefix_cache
+- model_type:
+  - chat
+  - vl
+  engine_config:
+    tp: 2
+    dp: 4
+    ep: 8
+    extra:
+      max-batch-size: 256
+      cache-max-entry-count: 0.7
+      prefix-cache-decode-state-interval: 1024
+      prefix-cache-state-budget: 256
+  backends:
   - name: pytorch
     communicators:
     - nccl
   test_coverage:
-  - prefix_cache
-  quantization:
-    pytorch:
-    - fp8
+  - prefix_cache_evaluate
 - model_type:
   - chat
   - vl
@@ -105,7 +117,6 @@ h:
   - longtext_benchmark
   - longtext_evaluate
   - mllm_evaluate
-  - prefix_cache
 ascend:
 - model_type:
   - chat

autotest/configs/Qwen/Qwen3.5-4B.yml

@@ -15,3 +15,17 @@ h:
     - nccl
   test_coverage:
   - func
+- model_type:
+  - chat
+  - vl
+  engine_config:
+    tp: 1
+    extra:
+      prefix-cache-state-budget: 256
+      max-prefill-token-num: 64
+  backends:
+  - name: pytorch
+    communicators:
+    - nccl
+  test_coverage:
+  - prefix_cache

autotest/configs/README.md

@@ -96,12 +96,16 @@ Available keys:
 - `longtext_evaluate`
 - `mllm_evaluate`
 - `prefix_cache`
+- `prefix_cache_evaluate`
 - `quantization`
 
 Rules:
 
 - Keep MTP (speculative decoding) on its own row with `speculative-algorithm` in `engine_config.extra`; use `func` and/or `evaluate` in `test_coverage`.
-- Use `prefix_cache` in `test_coverage`; do not add `enable-prefix-caching` manually to `engine_config.extra`.
+- Use `prefix_cache` / `prefix_cache_evaluate` in `test_coverage`; do not add `enable-prefix-caching` manually to `engine_config.extra`.
+- SSM models (Qwen3.5 / Intern-S2): **separate** dedicated rows, **pytorch only** (TurboMind has no serve CLI for checkpoint interval; do not list turbomind on this row). Do not put these knobs on shared `func`/`evaluate` rows. Do not inject extras in Python.
+  - **Func** (`tools/restful` / pipeline): `test_coverage: [prefix_cache]`; `max-prefill-token-num: 64` + `prefix-cache-state-budget`.
+  - **Accuracy** (`evaluate` `*_prefix_cache_*` / `submit_prefix_accuracy_job`): own row with `test_coverage: [prefix_cache_evaluate]`; `prefix-cache-state-budget` + `prefix-cache-decode-state-interval`（**不要** `max-prefill-token-num: 64`）.
 - Use `quantization` in `test_coverage` only for runtime weight-quant rows (`awq`, `gptq`, `w8a8`).
 
 ## `interface` (REST interface coverage)
@@ -189,10 +193,29 @@ is enabled.
 Suites:
 
 - `base` — chat/completions basic cases; generate without logprob/experts
+
 - `logprob` — generate logprob cases
+
 - `experts` — generate routed-experts cases
 - `anthropic` — Anthropic Messages HTTP + SDK smoke (chat protocol models from yaml). Share a profile with `base`/`logprob` when `extra` matches; otherwise its own profile **without** `tool-call-parser` / `reasoning-parser`.
 - `toolcall` — `interface/restful/tool_parser/` (requires `tool-call-parser` in yaml `extra`; add `enable-return-routed-experts: true` when toolcall includes `@experts` cases)
+
+- `hard_schema` — walle MFJS tool-call schema validation (`tool_parser/test_tool_call_json_schema.py`; requires `tool-call-parser`; for thinking models set `gen_config.chat-template-kwargs.enable_thinking: true` (not `interface.extra`); runs full walle case set). Kimi-Vendor-Verifier checkout: `eval_resource/Kimi-Vendor-Verifier` (see `kimi_vendor_verifier_path` in `env_paths.yml`), overridable via `KIMI_VENDOR_VERIFIER_ROOT`; cases override via `WALLE_CASE_DIR`. One representative model per parser (see table below); all `internlm/*` configs with `toolcall` also enable `hard_schema`.
+
+  | `tool-call-parser` | Representative model                                    |
+  | ------------------ | ------------------------------------------------------- |
+  | `qwen3coder`       | `Qwen/Qwen3.5-35B-A3B`                                  |
+  | `qwen3`            | `Qwen/Qwen3-8B-FP8`                                     |
+  | `qwen2d5`          | `Qwen/Qwen2.5-7B-Instruct`                              |
+  | `llama3`           | `meta-llama/Llama-3.1-70B-Instruct`                     |
+  | `glm47`            | `zai-org/GLM-4.7-Flash`                                 |
+  | `kimi-k2`          | `moonshotai/Kimi-K2-Instruct-0905`                      |
+  | `deepseek-v32`     | `deepseek-ai/DeepSeek-V3`                               |
+  | `deepseek-v4`      | `deepseek-ai/DeepSeek-V4-Flash-0731`                    |
+  | `intern-s1`        | `internlm/Intern-S1` (+ all Intern-S1 / Intern-S1-mini) |
+  | `interns2-preview` | `internlm/Intern-S2-Preview` (+ all Intern-S2 variants) |
+  | `internlm`         | `internlm/internlm3-8b-instruct`                        |
+
 - `reasoning` — `interface/restful/reasoning_parser/` (requires `reasoning-parser` in yaml `extra`)
 
 Notes:

autotest/configs/deepseek-ai/DeepSeek-V3.yml

@@ -10,6 +10,18 @@ h:
     - nccl
   test_coverage:
   - func
+  interface:
+    pytorch:
+    - suites:
+      - base
+      - logprob
+      - toolcall
+      - hard_schema
+      extra:
+        logprobs-mode: raw_logprobs
+        enable-return-routed-experts: true
+        tool-call-parser: deepseek-v32
+        reasoning-parser: deepseek-v3
 - model_type: chat
   engine_config:
     tp: 16