[Sarvam] Add inference support for Sarvam MoE LLMs (#18938)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

docs/supported_models/text_generation/generative_models.md

@@ -69,3 +69,4 @@ in the GitHub search bar.
 | **Falcon-H1** (0.5B–34B) | `tiiuae/Falcon-H1-34B-Instruct` | TII's hybrid Mamba-Transformer architecture combining attention and state-space models for efficient long-context inference. |
 | **Hunyuan-Large** (389B, MoE) | `tencent/Tencent-Hunyuan-Large` | Tencent's open-source MoE model with 389B total / 52B active parameters, featuring Cross-Layer Attention (CLA) for improved efficiency. |
 | **IBM Granite 4.0 (Hybrid, Dense)** | `ibm-granite/granite-4.0-h-micro`, `ibm-granite/granite-4.0-micro` | IBM Granite 4.0 micro models: hybrid Mamba–MoE (`h-micro`) and dense (`micro`) variants. Enterprise-focused reasoning models |
+| **Sarvam 2** (30B-A2B, 105B-A10B) | `sarvamai/sarvam-2` | Sarvam's Mixture-of-Experts models. The 105B variant uses MLA (Multi-head Latent Attention) and the 30B variant uses GQA, both with 128 routed experts. |

python/sglang/srt/configs/model_config.py

@@ -506,6 +506,23 @@ class ModelConfig:
                 scaling_factor = self.hf_config.rope_scaling["factor"]
                 mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
                 self.scaling = self.scaling * mscale * mscale
+        elif "SarvamMLAForCausalLM" in self.hf_config.architectures:
+            self.head_dim = (
+                self.hf_config.qk_nope_head_dim + self.hf_config.qk_rope_head_dim
+            )
+            self.attention_arch = AttentionArch.MLA
+            self.kv_lora_rank = self.hf_config.kv_lora_rank
+            self.qk_rope_head_dim = self.hf_config.qk_rope_head_dim
+            self.qk_nope_head_dim = self.hf_config.qk_nope_head_dim
+            self.v_head_dim = self.hf_config.v_head_dim
+            self.scaling = 1 / math.sqrt(self.qk_nope_head_dim + self.qk_rope_head_dim)
+            if self.hf_config.rope_scaling:
+                mscale_all_dim = self.hf_config.rope_scaling.get(
+                    "mscale_all_dim", False
+                )
+                scaling_factor = self.hf_config.rope_scaling["factor"]
+                mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
+                self.scaling = self.scaling * mscale * mscale
         else:
             if (
                 "MistralModel" in self.hf_config.architectures