[V1][Mamba] - Enable V1 by default for Mamba Models

tests/models/language/generation/test_hybrid.py

@@ -100,21 +100,21 @@ def test_models(
         else:
             hf_outputs = None
 
-    if model not in V0_UNSUPPORTED_MODELS:
-        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
-            vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
-                example_prompts, max_tokens, num_logprobs)
-    else:
-        vllm_v0_outputs = None
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "0")
+        if model not in V0_UNSUPPORTED_MODELS:
+            with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
+                vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
+                    example_prompts, max_tokens, num_logprobs)
+        else:
+            vllm_v0_outputs = None
 
     if model in V1_SUPPORTED_MODELS:
-        with monkeypatch.context() as m:
-            m.setenv("VLLM_USE_V1", "1")
-            with vllm_runner(model,
-                             max_num_seqs=MAX_NUM_SEQS,
-                             enable_prefix_caching=False) as vllm_model:
-                vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
-                    example_prompts, max_tokens, num_logprobs)
+        with vllm_runner(model,
+                         max_num_seqs=MAX_NUM_SEQS,
+                         enable_prefix_caching=False) as vllm_model:
+            vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
+                example_prompts, max_tokens, num_logprobs)
     else:
         vllm_v1_outputs = None
 
@@ -137,7 +137,7 @@ def test_models(
         )
 
 
-@pytest.mark.parametrize("model", SSM_MODELS + HYBRID_MODELS)
+@pytest.mark.parametrize("model", SSM_MODELS[0] + HYBRID_MODELS[0])
 @pytest.mark.parametrize("max_tokens", [64])
 @pytest.mark.parametrize("num_logprobs", [5])
 def test_batching(
@@ -402,24 +402,21 @@ def test_full_cuda_graph(
         else:
             hf_outputs = None
 
-    if model not in V0_UNSUPPORTED_MODELS:
-        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
-            vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
-                example_prompts, max_tokens, num_logprobs)
-    else:
-        vllm_v0_outputs = None
-
     with monkeypatch.context() as m:
-        m.setenv("VLLM_USE_V1", "1")
-        if model in HYBRID_MODELS:
-            # required due to reorder_batch behaviour
-            m.setenv("VLLM_ATTENTION_BACKEND", "FLASHINFER")
-        with vllm_runner(model,
-                         max_num_seqs=MAX_NUM_SEQS,
-                         compilation_config={'full_cuda_graph': True},
-                         enable_prefix_caching=False) as vllm_model:
-            vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
-                example_prompts, max_tokens, num_logprobs)
+        m.setenv("VLLM_USE_V1", "0")
+        if model not in V0_UNSUPPORTED_MODELS:
+            with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
+                vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
+                    example_prompts, max_tokens, num_logprobs)
+        else:
+            vllm_v0_outputs = None
+
+    with vllm_runner(model,
+                     max_num_seqs=MAX_NUM_SEQS,
+                     compilation_config={'full_cuda_graph': True},
+                     enable_prefix_caching=False) as vllm_model:
+        vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
 
     if hf_outputs is not None and vllm_v0_outputs is not None:
         check_logprobs_close(
@@ -466,24 +463,21 @@ def test_fp32_state(
         else:
             hf_outputs = None
 
-    with vllm_runner(model,
-                     max_num_seqs=MAX_NUM_SEQS,
-                     mamba_ssm_cache_dtype="float32") as vllm_model:
-        vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
-            example_prompts, max_tokens, num_logprobs)
-
     with monkeypatch.context() as m:
-        m.setenv("VLLM_USE_V1", "1")
-        if model in HYBRID_MODELS:
-            # required due to reorder_batch behaviour
-            m.setenv("VLLM_ATTENTION_BACKEND", "FLASHINFER")
+        m.setenv("VLLM_USE_V1", "0")
         with vllm_runner(model,
                          max_num_seqs=MAX_NUM_SEQS,
-                         mamba_ssm_cache_dtype="float32",
-                         enable_prefix_caching=False) as vllm_model:
-            vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
+                         mamba_ssm_cache_dtype="float32") as vllm_model:
+            vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
                 example_prompts, max_tokens, num_logprobs)
 
+    with vllm_runner(model,
+                     max_num_seqs=MAX_NUM_SEQS,
+                     mamba_ssm_cache_dtype="float32",
+                     enable_prefix_caching=False) as vllm_model:
+        vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
     if hf_outputs is not None:
         check_logprobs_close(
             outputs_0_lst=hf_outputs,

vllm/engine/arg_utils.py

@@ -1452,11 +1452,6 @@ def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
                                recommend_to_remove=False)
             return False
 
-        # V1 mamba models are unoptimized.
-        if model_config.has_inner_state and _warn_or_fallback(
-                feature_name="Mamba"):
-            return False
-
         # No Concurrent Partial Prefills so far.
         if (self.max_num_partial_prefills
                 != SchedulerConfig.max_num_partial_prefills

vllm/model_executor/models/config.py

@@ -417,4 +417,5 @@ def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
     "GptOssForCausalLM": GptOssForCausalLMConfig,
     "MambaForCausalLM": MambaModelConfig,
     "Mamba2ForCausalLM": MambaModelConfig,
+    "FalconMambaForCausalLM": MambaModelConfig,
 }