[Model] Support Mamba (#6484)

.buildkite/run-cpu-test-ppc64le.sh

@@ -18,7 +18,13 @@ docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/hugg
 # Run basic model test
 docker exec cpu-test bash -c "
   pip install pytest matplotlib einops transformers_stream_generator
-  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_oot_registration.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py --ignore=tests/models/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
+  pytest -v -s tests/models -m \"not vlm\" \
+    --ignore=tests/models/test_embedding.py \
+    --ignore=tests/models/test_oot_registration.py \
+    --ignore=tests/models/test_registry.py \
+    --ignore=tests/models/test_jamba.py \
+    --ignore=tests/models/test_mamba.py \
+    --ignore=tests/models/test_danube3_4b.py" # Mamba kernels and Danube3-4B on CPU is not supported
 
 # online inference
 docker exec cpu-test bash -c "

.buildkite/run-cpu-test.sh

@@ -27,6 +27,7 @@ docker exec cpu-test bash -c "
   pytest -v -s tests/models/decoder_only/language \
     --ignore=tests/models/test_fp8.py \
     --ignore=tests/models/decoder_only/language/test_jamba.py \
+    --ignore=tests/models/decoder_only/language/test_mamba.py \
     --ignore=tests/models/decoder_only/language/test_granitemoe.py \
     --ignore=tests/models/decoder_only/language/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
 

docs/source/models/supported_models.rst

@@ -152,6 +152,11 @@ Text Generation
     - :code:`meta-llama/Meta-Llama-3.1-405B-Instruct`, :code:`meta-llama/Meta-Llama-3.1-70B`, :code:`meta-llama/Meta-Llama-3-70B-Instruct`, :code:`meta-llama/Llama-2-70b-hf`, :code:`01-ai/Yi-34B`, etc.
     - ✅︎
     - ✅︎
+  * - :code:`MambaForCausalLM`
+    - Mamba
+    - :code:`state-spaces/mamba-130m-hf`, :code:`state-spaces/mamba-790m-hf`, :code:`state-spaces/mamba-2.8b-hf`, etc.
+    - ✅︎
+    -
   * - :code:`MiniCPMForCausalLM`
     - MiniCPM
     - :code:`openbmb/MiniCPM-2B-sft-bf16`, :code:`openbmb/MiniCPM-2B-dpo-bf16`, etc.

tests/kernels/test_attention_selector.py

@@ -20,22 +20,22 @@ def test_env(name: str, device: str, monkeypatch):
 
     if device == "cpu":
         with patch("vllm.attention.selector.is_cpu", return_value=True):
-            backend = which_attn_to_use(8, 16, 8, None, torch.float16,
-                                        torch.float16, 16)
+            backend = which_attn_to_use(16, None, torch.float16, torch.float16,
+                                        16, False)
         assert backend.name == "TORCH_SDPA"
     elif device == "hip":
         with patch("vllm.attention.selector.is_hip", return_value=True):
-            backend = which_attn_to_use(8, 16, 8, None, torch.float16,
-                                        torch.float16, 16)
+            backend = which_attn_to_use(16, None, torch.float16, torch.float16,
+                                        16, False)
         assert backend.name == "ROCM_FLASH"
     elif device == "openvino":
         with patch("vllm.attention.selector.is_openvino", return_value=True):
-            backend = which_attn_to_use(8, 16, 8, None, torch.float16,
-                                        torch.float16, 16)
+            backend = which_attn_to_use(16, None, torch.float16, torch.float16,
+                                        16, False)
         assert backend.name == "OPENVINO"
     else:
-        backend = which_attn_to_use(8, 16, 8, None, torch.float16,
-                                    torch.float16, 16)
+        backend = which_attn_to_use(16, None, torch.float16, torch.float16, 16,
+                                    False)
         assert backend.name == name
 
 
@@ -46,37 +46,42 @@ def test_flash_attn(monkeypatch):
 
     # Unsupported CUDA arch
     with patch("torch.cuda.get_device_capability", return_value=(7, 5)):
-        backend = which_attn_to_use(8, 16, 8, None, torch.float16, None, 16)
+        backend = which_attn_to_use(16, None, torch.float16, None, 16, False)
         assert backend.name != STR_FLASH_ATTN_VAL
 
     # Unsupported data type
-    backend = which_attn_to_use(8, 16, 8, None, torch.float8_e4m3fn, None, 16)
+    backend = which_attn_to_use(16, None, torch.float8_e4m3fn, None, 16, False)
     assert backend.name != STR_FLASH_ATTN_VAL
 
     # Unsupported kv cache data type
-    backend = which_attn_to_use(8, 16, 8, None, torch.float16, "fp8", 16)
+    backend = which_attn_to_use(16, None, torch.float16, "fp8", 16, False)
     assert backend.name != STR_FLASH_ATTN_VAL
 
     # Unsupported block size
-    backend = which_attn_to_use(8, 16, 8, None, torch.float16, None, 8)
+    backend = which_attn_to_use(16, None, torch.float16, None, 8, False)
     assert backend.name != STR_FLASH_ATTN_VAL
 
     # Unsupported sliding window
-    backend = which_attn_to_use(8, 16, 8, 1, torch.float16, None, 16)
+    backend = which_attn_to_use(16, 1, torch.float16, None, 16, False)
     assert backend.name != STR_FLASH_ATTN_VAL
 
     # flash-attn is not installed
     with patch.dict('sys.modules', {'vllm_flash_attn': None}):
-        backend = which_attn_to_use(8, 16, 8, None, torch.float16, None, 16)
+        backend = which_attn_to_use(16, None, torch.float16, None, 16, False)
         assert backend.name != STR_FLASH_ATTN_VAL
 
     # Unsupported head size
-    backend = which_attn_to_use(8, 17, 8, None, torch.float16, None, 16)
+    backend = which_attn_to_use(17, None, torch.float16, None, 16, False)
+    assert backend.name != STR_FLASH_ATTN_VAL
+
+    # Attention-free models should bypass env and use PlaceholderAttention
+    backend = which_attn_to_use(16, None, torch.float16, torch.float16, 16,
+                                True)
     assert backend.name != STR_FLASH_ATTN_VAL
 
 
 def test_invalid_env(monkeypatch):
     """Throw an exception if the backend name is invalid."""
     override_backend_env_variable(monkeypatch, STR_INVALID_VAL)
     with pytest.raises(ValueError):
-        which_attn_to_use(8, 16, 8, None, torch.float16, None, 16)
+        which_attn_to_use(16, None, torch.float16, None, 16, False)