Add Qwen 3.6 MoE quantization regressions (#2752)

Author: Qubitium

tests/models/test_qwen3_6_moe.py

@@ -0,0 +1,85 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import os
+
+from model_test import ModelTest
+
+from gptqmodel.quantization.config import ExpertsRoutingOverride, Fallback, MoEConfig, VramStrategy
+
+
+LAST_FOUR_ONLY_NEGATIVE_MATCH = r"^model\.language_model\.layers\.(?:[0-9]|[1-2][0-9]|3[0-5])(?:\.|$)"
+
+
+def _resolve_qwen3_6_moe_model_path() -> str:
+    """Prefer the user-requested local-dir path, but allow an explicit override for repro runs."""
+
+    override = os.environ.get("GPTQMODEL_QWEN3_6_MOE_MODEL_PATH")
+    if override:
+        return override
+
+    requested = "/moonster/data/model/Qwen3.6-35B-A3B"
+    legacy_fallback = "/monster/data/model/Qwen3.6-35B-A3B"
+    if os.path.isdir(requested):
+        return requested
+    if os.path.isdir(legacy_fallback):
+        return legacy_fallback
+    return requested
+
+
+class TestQwen3_6Moe(ModelTest):
+    """Fast 3-A100 compat regression for the Qwen 3.6 MoE checkpoint released on the Qwen hub."""
+
+    FALLBACK = Fallback()
+    NATIVE_MODEL_ID = _resolve_qwen3_6_moe_model_path()
+    DATASET_SIZE = 16
+    DATASET_CONCAT_SIZE = 1024
+    QUANT_BATCH_SIZE = 1
+    EVAL_BATCH_SIZE = "auto"
+    # Keep post-quant validation spread across the visible A100 pool instead of forcing a single-GPU reload.
+    EVAL_SINGLE_GPU = False
+    # The native checkpoint has 40 decoder layers; skip 0-35 so only the last four are quantized.
+    DYNAMIC = {
+        f"-:{LAST_FOUR_ONLY_NEGATIVE_MATCH}": {},
+    }
+    DENSE_VRAM_STRATEGY = VramStrategy.EXCLUSIVE
+    # Keep the dense shell on the first visible device and spread expert work across the rest.
+    DENSE_VRAM_STRATEGY_DEVICES = ["cuda:0"]
+    MOE_VRAM_STRATEGY = VramStrategy.BALANCED
+    # The mixed RTX 4090 slot on this host OOMs during late replay for this 35B MoE path,
+    # so keep the quantization pool on the A100-class devices.
+    MOE_VRAM_STRATEGY_DEVICES = ["cuda:1", "cuda:2"]
+    # Fast compat mode already trims the decoder stack; keep routing aligned with the checkpoint default.
+    MOE_CONFIG = MoEConfig(routing=ExpertsRoutingOverride(num_experts_per_tok=8))
+    EVAL_TASKS_FAST = {
+        "gsm8k_platinum_cot": {
+            "chat_template": False,
+            # Evaluate the already-loaded post-quant wrapper so the three-A100 placement survives into Evalution.
+            "evalution_batch_size": 4,
+            "evalution_model_args": {
+                "dtype": "bfloat16",
+                "attn_implementation": "paged|flash_attention_2",
+                "device_map": "auto",
+            },
+            "evalution_suite_kwargs": {
+                "batch_size": 4,
+                "max_new_tokens": 256,
+                "stream": True,
+                "max_rows": 8,
+            },
+            "acc,num": {
+                "value": {
+                    # Recorded on 2026-04-16 with three visible A100s (dense=cuda:0, MoE=cuda:1,cuda:2).
+                    "A100": 1.0,
+                },
+                "floor_pct": 0.04,
+                "ceil_pct": 1.0,
+            },
+        },
+    }
+    EVAL_TASKS_SLOW = EVAL_TASKS_FAST
+
+    def test_qwen3_6_moe(self):
+        self.quantize_and_evaluate()

tests/test_qwen3_6_moe_support.py

@@ -0,0 +1,76 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import json
+
+from transformers import AutoConfig
+
+from gptqmodel.models.auto import check_and_get_model_definition
+from gptqmodel.models.definitions.qwen3_5_moe import Qwen3_5_MoeQModel
+
+
+def test_qwen3_6_moe_reuses_the_qwen3_5_moe_transformers_definition(tmp_path):
+    """Guard the real Qwen 3.6 MoE config shape shipped on the Hub."""
+
+    layer_types = ["linear_attention", "linear_attention", "linear_attention", "full_attention"] * 10
+    config = {
+        "architectures": ["Qwen3_5MoeForConditionalGeneration"],
+        "image_token_id": 248056,
+        "model_type": "qwen3_5_moe",
+        "text_config": {
+            "dtype": "bfloat16",
+            "full_attention_interval": 4,
+            "hidden_size": 2048,
+            "layer_types": layer_types,
+            "max_position_embeddings": 262144,
+            "model_type": "qwen3_5_moe_text",
+            "moe_intermediate_size": 512,
+            "num_attention_heads": 16,
+            "num_experts": 256,
+            "num_experts_per_tok": 8,
+            "num_hidden_layers": 40,
+            "num_key_value_heads": 2,
+            "partial_rotary_factor": 0.25,
+            "rope_parameters": {
+                "mrope_interleaved": True,
+                "mrope_section": [11, 11, 10],
+                "partial_rotary_factor": 0.25,
+                "rope_theta": 10000000,
+                "rope_type": "default",
+            },
+            "shared_expert_intermediate_size": 512,
+            "tie_word_embeddings": False,
+            "use_cache": True,
+            "vocab_size": 248320,
+        },
+        "tie_word_embeddings": False,
+        "transformers_version": "4.57.1",
+        "video_token_id": 248057,
+        "vision_config": {
+            "deepstack_visual_indexes": [],
+            "depth": 27,
+            "hidden_size": 1152,
+            "in_channels": 3,
+            "intermediate_size": 4304,
+            "model_type": "qwen3_5_moe",
+            "num_heads": 16,
+            "num_position_embeddings": 2304,
+            "out_hidden_size": 2048,
+            "patch_size": 16,
+            "spatial_merge_size": 2,
+            "temporal_patch_size": 2,
+        },
+        "vision_end_token_id": 248054,
+        "vision_start_token_id": 248053,
+    }
+    model_dir = tmp_path / "qwen3_6_moe"
+    model_dir.mkdir()
+    (model_dir / "config.json").write_text(json.dumps(config), encoding="utf-8")
+
+    resolved_config = AutoConfig.from_pretrained(model_dir, trust_remote_code=False)
+    model_definition = check_and_get_model_definition(model_dir, trust_remote_code=False)
+
+    assert type(resolved_config).__name__ == "Qwen3_5MoeConfig"
+    assert model_definition is Qwen3_5_MoeQModel