[NPU] bugfix with Kimi-k2 and bge-reranker-v2 model (#17478)

Co-authored-by: amote-i <49533125+amote-i@users.noreply.github.com> Co-authored-by: cy <chenyang08056032@163.com>
2026-01-22 22:02:05 +08:00
parent a4dc432587
commit 5d299c25c0
5 changed files with 130 additions and 11 deletions
@@ -0,0 +1,112 @@
+import multiprocessing as mp
+import unittest
+from typing import Optional
+
+import torch
+from transformers import AutoConfig, AutoTokenizer
+
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.runners import HFRunner, SRTRunner
+from sglang.test.test_utils import CustomTestCase, get_similarities
+
+register_npu_ci(
+    est_time=400,
+    suite="nightly-1-npu-a3",
+    nightly=True,
+    disabled="embeddings are not all close",
+)
+
+DEFAULT_PROMPTS = [
+    "The capital of the United Kingdom is",
+    "Today is a sunny day and I like",
+    "AI is a field of computer science focused on",
+]
+
+MODELS = [
+    ("/root/.cache/modelscope/hub/models/bge-large-en-v1.5", 1, 1e-5),
+]
+TORCH_DTYPES = [torch.float16]
+
+
+class TestEmbeddingModels(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        mp.set_start_method("spawn", force=True)
+
+    def _truncate_prompts(self, prompts, model_path):
+        config = AutoConfig.from_pretrained(model_path)
+        max_length = getattr(config, "max_position_embeddings", 2048)
+
+        tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+        truncated_prompts = []
+        for prompt in prompts:
+            tokens = tokenizer(prompt, return_tensors="pt", truncation=False)
+            if len(tokens.input_ids[0]) > max_length:
+                truncated_text = tokenizer.decode(
+                    tokens.input_ids[0][: max_length - 1], skip_special_tokens=True
+                )
+                truncated_prompts.append(truncated_text)
+            else:
+                truncated_prompts.append(prompt)
+        return truncated_prompts
+
+    def assert_close_prefill_logits(
+        self,
+        prompts,
+        model_path,
+        tp_size,
+        torch_dtype,
+        prefill_tolerance,
+        matryoshka_dim: Optional[int] = None,
+    ) -> None:
+        truncated_prompts = self._truncate_prompts(prompts, model_path)
+
+        with HFRunner(
+            model_path,
+            torch_dtype=torch_dtype,
+            model_type="embedding",
+            matryoshka_dim=matryoshka_dim,
+        ) as hf_runner:
+            hf_outputs = hf_runner.forward(truncated_prompts)
+
+        attention_backend = "ascend"
+        with SRTRunner(
+            model_path,
+            tp_size=tp_size,
+            torch_dtype=torch_dtype,
+            model_type="embedding",
+            attention_backend=attention_backend,
+            json_model_override_args=(
+                {"matryoshka_dimensions": [matryoshka_dim]} if matryoshka_dim else None
+            ),
+        ) as srt_runner:
+            srt_outputs = srt_runner.forward(
+                truncated_prompts, dimensions=matryoshka_dim
+            )
+
+        for i in range(len(prompts)):
+            hf_logits = torch.Tensor(hf_outputs.embed_logits[i])
+            srt_logits = torch.Tensor(srt_outputs.embed_logits[i])
+
+            similarity = torch.tensor(get_similarities(hf_logits, srt_logits))
+            print("similarity diff", abs(similarity - 1))
+
+            if len(prompts[i]) <= 1000:
+                assert torch.all(
+                    abs(similarity - 1) < prefill_tolerance
+                ), "embeddings are not all close"
+
+    def test_prefill_logits(self):
+        models_to_test = MODELS
+
+        for model, tp_size, prefill_tolerance in models_to_test:
+            for torch_dtype in TORCH_DTYPES:
+                self.assert_close_prefill_logits(
+                    DEFAULT_PROMPTS, model, tp_size, torch_dtype, prefill_tolerance
+                )
+
+
+if __name__ == "__main__":
+    unittest.main()