[Intel][XPU][LoRA] Enable LoRA on Intel XPU (#30345)

python/sglang/test/lora_utils.py

@@ -4,9 +4,36 @@ from typing import List, Optional
 
 import torch
 
+from sglang.srt.utils import is_xpu
 from sglang.test.runners import HFRunner, SRTRunner
 from sglang.test.test_utils import calculate_rouge_l
 
+_IS_XPU = is_xpu()
+
+
+def _assert_lora_output_match(
+    srt_str: str, hf_str: str, rouge_tol: float, context: str
+):
+    """Compare SRT vs HF greedy output strings.
+
+    Everywhere except XPU we keep the historical strict exact-match (SGLang and HF
+    kernels agree numerically enough for greedy argmax to pick identical tokens).
+    On XPU, small kernel-level fp differences can make greedy decoding diverge
+    after a shared prefix even when the LoRA math is correct, so we fall back to
+    the same ROUGE-L tolerance the per-adaptor comparison path uses.
+    """
+    srt_str = srt_str.strip(" ")
+    hf_str = hf_str.strip(" ")
+    if not _IS_XPU:
+        assert srt_str == hf_str, (srt_str, hf_str)
+        return
+    rouge_score = calculate_rouge_l([srt_str], [hf_str])[0]
+    if rouge_score < rouge_tol:
+        raise AssertionError(
+            f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} for {context}. "
+            f"SRT: {srt_str!r} HF: {hf_str!r}"
+        )
+
 
 @dataclasses.dataclass
 class LoRAAdaptor:
@@ -624,17 +651,22 @@ def run_lora_test_by_batch(
         print("HF output:", hf_output_str)
         print("SRT no lora output:", srt_no_lora_outputs.output_strs[i].strip())
         print("HF no lora output:", hf_no_lora_outputs.output_strs[i].strip())
-        assert srt_outputs.output_strs[i].strip(" ") == hf_outputs.output_strs[i].strip(
-            " "
-        ), (
-            srt_outputs.output_strs[i].strip(" "),
-            hf_outputs.output_strs[i].strip(" "),
+        rouge_tol = (
+            adaptors[i].rouge_l_tolerance
+            if adaptors[i].rouge_l_tolerance is not None
+            else model_case.rouge_l_tolerance
         )
-        assert srt_no_lora_outputs.output_strs[i].strip(
-            " "
-        ) == hf_no_lora_outputs.output_strs[i].strip(" "), (
-            srt_no_lora_outputs.output_strs[i].strip(" "),
-            hf_no_lora_outputs.output_strs[i].strip(" "),
+        _assert_lora_output_match(
+            srt_outputs.output_strs[i],
+            hf_outputs.output_strs[i],
+            rouge_tol,
+            f"base '{base_path}', adaptor '{adaptor_names[i]}', backend '{backend}' (LoRA)",
+        )
+        _assert_lora_output_match(
+            srt_no_lora_outputs.output_strs[i],
+            hf_no_lora_outputs.output_strs[i],
+            rouge_tol,
+            f"base '{base_path}', backend '{backend}' (no-LoRA baseline)",
         )