[Transformers v5] Add SarvamMLAConfig to fix SarvamMLAForCausalLM (#38734)

vllm/transformers_utils/config.py

@@ -116,6 +116,7 @@ def __getitem__(self, key):
     qwen3_5="Qwen3_5Config",
     qwen3_5_moe="Qwen3_5MoeConfig",
     lfm2_moe="Lfm2MoeConfig",
+    sarvam_mla="SarvamMLAConfig",
     tarsier2="Tarsier2Config",
 )
 

vllm/transformers_utils/configs/__init__.py

@@ -66,6 +66,7 @@
     "Qwen3_5TextConfig": "vllm.transformers_utils.configs.qwen3_5",
     "Qwen3_5MoeConfig": "vllm.transformers_utils.configs.qwen3_5_moe",
     "Qwen3_5MoeTextConfig": "vllm.transformers_utils.configs.qwen3_5_moe",
+    "SarvamMLAConfig": "vllm.transformers_utils.configs.sarvam",
     "Tarsier2Config": "vllm.transformers_utils.configs.tarsier2",
     # Special case: DeepseekV3Config is from HuggingFace Transformers
     "DeepseekV3Config": "transformers",
@@ -121,6 +122,7 @@
     "Qwen3_5TextConfig",
     "Qwen3_5MoeConfig",
     "Qwen3_5MoeTextConfig",
+    "SarvamMLAConfig",
     "Tarsier2Config",
 ]
 

vllm/transformers_utils/configs/sarvam.py

@@ -0,0 +1,114 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""SarvamMLA model configuration.
+
+This vLLM-local config replaces the trust_remote_code configuration shipped
+with sarvamai/sarvam-105b.  The remote configuration calls
+``validate_rope(ignore_keys=...)`` which was broken by the transformers v5
+API change that removed the ``ignore_keys`` parameter (upstream PR #41250).
+
+By registering ``SarvamMLAConfig`` in vLLM's ``_CLASS_TO_MODULE`` table,
+``HFConfigParser.parse()`` loads *this* class instead of the remote one,
+bypassing ``trust_remote_code`` and the broken API call entirely.
+"""
+
+from transformers import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class SarvamMLAConfig(PretrainedConfig):
+    """Configuration for SarvamMLA models (e.g., ``sarvamai/sarvam-105b``).
+
+    Mirrors the fields present in the model's ``config.json`` and the
+    attributes accessed by ``vllm/model_executor/models/sarvam.py``.
+    """
+
+    model_type = "sarvam_mla"
+
+    def __init__(
+        self,
+        vocab_size: int = 262144,
+        hidden_size: int = 4096,
+        intermediate_size: int = 16384,
+        moe_intermediate_size: int | None = None,
+        num_hidden_layers: int = 32,
+        num_attention_heads: int = 64,
+        num_key_value_heads: int | None = None,
+        kv_lora_rank: int = 512,
+        qk_nope_head_dim: int = 128,
+        qk_rope_head_dim: int = 64,
+        v_head_dim: int = 128,
+        q_lora_rank: int | None = None,
+        max_position_embeddings: int = 131072,
+        rms_norm_eps: float = 1e-6,
+        hidden_act: str = "silu",
+        num_experts: int | None = None,
+        num_experts_per_tok: int | None = None,
+        moe_shared_expert_intermediate_size: int | None = None,
+        first_k_dense_replace: int = 1,
+        rope_theta: float = 10000.0,
+        rope_parameters: dict | None = None,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = False,
+        **kwargs,
+    ):
+        # Handle rope_scaling → rope_parameters conversion so that both the
+        # old-style {"type": ...} field and the v5 {"rope_type": ...} field
+        # are accepted.
+        rope_scaling = kwargs.pop("rope_scaling", None)
+        if rope_scaling is not None and rope_parameters is None:
+            rope_parameters = dict(rope_scaling)
+            # Normalise "type" → "rope_type" (transformers v5 convention).
+            if "type" in rope_parameters and "rope_type" not in rope_parameters:
+                rope_parameters["rope_type"] = rope_parameters.pop("type")
+        if rope_parameters is None:
+            rope_parameters = {"rope_type": "default"}
+        if "rope_theta" not in rope_parameters:
+            rope_parameters["rope_theta"] = rope_theta
+        self.rope_parameters = rope_parameters
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        # ``sarvam.py`` accesses ``config.moe_intermediate_size`` directly.
+        # The config.json stores a single ``intermediate_size`` field; expose
+        # it under both names so the model layer and dense-FFN path both work.
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = (
+            moe_intermediate_size
+            if moe_intermediate_size is not None
+            else intermediate_size
+        )
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = (
+            num_key_value_heads
+            if num_key_value_heads is not None
+            else num_attention_heads
+        )
+
+        # MLA-specific attention dimensions.
+        self.kv_lora_rank = kv_lora_rank
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.q_lora_rank = q_lora_rank
+
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+        self.hidden_act = hidden_act
+        self.use_cache = use_cache
+
+        # MoE topology.
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_shared_expert_intermediate_size = moe_shared_expert_intermediate_size
+        self.first_k_dense_replace = first_k_dense_replace
+
+        super().__init__(
+            use_cache=use_cache,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )