[sgl-router] Match DeepSeek V4 rendering to SGLang (#40530)

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-09-21 17:56:41 +08:00
parent 2016f5e7a1
commit 0abb251a20
8 changed files with 526 additions and 51 deletions
@@ -0,0 +1,93 @@
+"""Regenerate native DeepSeek fixtures with SGLang's actual serving pipeline.
+
+Run from experimental/sgl-router in a SGLang Python environment:
+    python tests/scripts/generate_deepseek_parity.py
+
+The Rust tests always compare rendered text, even without cached model files.
+With the pinned HF snapshot cached they also compare exact token-ID digests.
+"""
+
+import copy
+import hashlib
+import json
+import os
+from pathlib import Path
+from types import SimpleNamespace
+
+from huggingface_hub import snapshot_download
+from transformers import AutoTokenizer
+
+os.environ["SGLANG_DEFAULT_THINKING"] = "false"
+os.environ["SGLANG_DSV4_REASONING_EFFORT"] = ""
+
+from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
+from sglang.srt.entrypoints.openai.serving_chat import OpenAIServingChat
+
+ROOT = Path(__file__).resolve().parents[1] / "fixtures/deepseek"
+MODELS = {
+    "v4": ("deepseek-ai/DeepSeek-V4-Flash", "60d8d70770c6776ff598c94bb586a859a38244f1"),
+}
+
+
+class RecordingTokenizer:
+    def __init__(self, tokenizer):
+        self.tokenizer = tokenizer
+        self.texts = []
+
+    def encode(self, text):
+        self.texts.append(text)
+        return self.tokenizer.encode(text)
+
+    def __getattr__(self, name):
+        return getattr(self.tokenizer, name)
+
+
+def main():
+    ROOT.mkdir(exist_ok=True)
+    for family, (model, revision) in MODELS.items():
+        path = snapshot_download(model, revision=revision, local_files_only=True)
+        tok = RecordingTokenizer(
+            AutoTokenizer.from_pretrained(path, local_files_only=True)
+        )
+        server = object.__new__(OpenAIServingChat)
+        server.chat_encoding_spec = "ds" + family
+        server.tokenizer_manager = SimpleNamespace(tokenizer=tok)
+        server.template_manager = SimpleNamespace(
+            jinja_template_content_format="string"
+        )
+        fixture_path = ROOT / (family + ".json")
+        fixture = json.loads(fixture_path.read_text())
+        for case in fixture["cases"]:
+            server._dsv4_reasoning_effort_profile = case["profile"]
+            tok.texts.clear()
+            request = ChatCompletionRequest(**copy.deepcopy(case["request"]))
+            # _convert_to_internal_request: kwargs effort replaces the request effort.
+            if request.chat_template_kwargs:
+                effort = request.chat_template_kwargs.pop("reasoning_effort", None)
+                if effort is not None:
+                    request.reasoning_effort = effort
+            ids = server._apply_jinja_template(
+                request, tools=None, is_multimodal=False
+            ).prompt_ids
+            case.update(
+                prompt="".join(tok.texts),
+                token_count=len(ids),
+                token_sha256=hashlib.sha256(
+                    b"".join(i.to_bytes(4, "little") for i in ids)
+                ).hexdigest(),
+            )
+        metadata = {k: v for k, v in fixture.items() if k != "cases"}
+        fixture_path.write_text(
+            json.dumps(metadata)[:-1]
+            + ', "cases": [\n'
+            + ",\n".join(
+                json.dumps(c, ensure_ascii=False, separators=(",", ":"))
+                for c in fixture["cases"]
+            )
+            + "\n]}\n"
+        )
+        print(f"wrote {family}: {len(fixture['cases'])} cases")
+
+
+if __name__ == "__main__":
+    main()