Add Deepseek mapping for RL.

abhinavclemson · abhinavclemson · commit 3bceb75e3273 · 2026-01-22T21:38:46.000Z
diff --git a/src/MaxText/integration/tunix/weight_mapping/__init__.py b/src/MaxText/integration/tunix/weight_mapping/__init__.py
@@ -18,7 +18,8 @@
 dispatcher to retrieve the correct weight mapping configuration for a given
 model name. This allows for easy extension to support new models.
 """
-
+from MaxText.integration.tunix.weight_mapping.deepseek3 import DEEPSEEK_VLLM_MAPPING
+from MaxText.integration.tunix.weight_mapping.gpt_oss import GPT_OSS_VLLM_MAPPING
 from MaxText.integration.tunix.weight_mapping.llama3 import LLAMA3_VLLM_MAPPING
 from MaxText.integration.tunix.weight_mapping.qwen3 import QWEN3_VLLM_MAPPING
 
@@ -31,6 +32,10 @@ def __getattr__(self, name):
       return LLAMA3_VLLM_MAPPING
     elif name.startswith("qwen3"):
       return QWEN3_VLLM_MAPPING
+    elif name.startswith("deepseek3"):
+      return DEEPSEEK_VLLM_MAPPING
+    elif name.startswith("gpt-oss"):
+      return GPT_OSS_VLLM_MAPPING
     else:
       raise ValueError(f"{name} vLLM weight mapping not found.")
 
diff --git a/src/MaxText/integration/tunix/weight_mapping/deepseek3.py b/src/MaxText/integration/tunix/weight_mapping/deepseek3.py
@@ -0,0 +1,157 @@
+# Copyright 2023–2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Mapping MaxText Deepseek (MoE) weights to vLLM/tpu-inference keys."""
+
+from dataclasses import dataclass
+
+
+@dataclass
+class DEEPSEEK_VLLM_MAPPING:
+  """Mapping MaxText Deepseek-V3 weights to Tunix/vLLM NNX keys."""
+
+  @staticmethod
+  def to_hf_hook_fns():
+    def flatten_3d_to_2d(val):
+      # Converts (Rank, Heads, HeadDim) -> (Rank, Heads * HeadDim)
+      if val.ndim == 3:
+        return val.reshape(val.shape[0], -1)
+      return val
+
+    return {
+        # MaxText MLA weights are 3D (Rank, Heads, HeadDim).
+        # tpu-inference expects 2D (Rank, Heads*HeadDim) before it splits them.
+        "base.decoder.layers.self_attention.wq_b.kernel": flatten_3d_to_2d,
+        "base.decoder.layers.self_attention.wkv_b.kernel": flatten_3d_to_2d,
+        "base.decoder.layers.self_attention.out.kernel": flatten_3d_to_2d,
+    }
+
+  @staticmethod
+  def to_hf_transpose_keys():
+    """Returns a list of keys for weights that need to be transposed.
+
+    Returns:
+      An empty dictionary, as no keys require transposition for this mapping.
+    """
+    return {}
+
+  @staticmethod
+  def lora_to_hf_mappings():
+    """Provides the mapping for LoRA (Low-Rank Adaptation) weights.
+
+    Returns:
+      None, as LoRA mappings are not defined for this model.
+    """
+    return None
+
+  @staticmethod
+  def to_hf_mapping():
+    mapping = {
+        # --- Base Model Params ---
+        # Map to HF names to be safe with loader regexes
+        "base.token_embedder.embedding": ("model.embed_tokens.weight", ("model", None)),
+        "base.decoder.decoder_norm.scale": ("model.norm.weight", (None,)),
+        "base.decoder.logits_dense.kernel": ("lm_head.weight", (None, "model")),
+        # MLA LAYERS (Map to HF Keys to trigger loader splitting logic)
+        # Norms
+        "base.decoder.layers.pre_self_attention_layer_norm.scale": (
+            "model.layers.*.input_layernorm.weight",
+            (None, "layer"),
+        ),
+        "base.decoder.layers.post_self_attention_layer_norm.scale": (
+            "model.layers.*.post_attention_layernorm.weight",
+            (None, "layer"),
+        ),
+        # MLA Norms
+        "base.decoder.layers.self_attention.kv_norm.scale": (
+            "model.layers.*.self_attn.kv_a_layernorm.weight",
+            (None, "layer"),
+        ),
+        "base.decoder.layers.self_attention.q_norm.scale": (
+            "model.layers.*.self_attn.q_a_layernorm.weight",
+            (None, "layer"),
+        ),
+        # MLA Projections
+        # We use HF names here so `DeepSeekV3WeightLoader` detects "kv_b_proj"
+        # and performs the necessary split into k_b and v_b for the MLA kernel.
+        "base.decoder.layers.self_attention.wq_a.kernel": (
+            "model.layers.*.self_attn.q_a_proj.weight",
+            (None, "layer", "model", None),
+        ),
+        "base.decoder.layers.self_attention.wq_b.kernel": (
+            "model.layers.*.self_attn.q_b_proj.weight",
+            (None, "layer", "model", None),
+        ),
+        "base.decoder.layers.self_attention.wkv_a.kernel": (
+            "model.layers.*.self_attn.kv_a_proj_with_mqa.weight",
+            (None, "layer", "model", None),
+        ),
+        "base.decoder.layers.self_attention.wkv_b.kernel": (
+            "model.layers.*.self_attn.kv_b_proj.weight",
+            (None, "layer", "model", None),
+        ),
+        "base.decoder.layers.self_attention.out.kernel": (
+            "model.layers.*.self_attn.o_proj.weight",
+            ("model", "layer", None, None),
+        ),
+        # DENSE MLP LAYERS (Map to vllm keys for safety/consistency)
+        "base.decoder.layers.mlp.wi_0.kernel": ("model.layers.*.mlp.gate_proj.weight", (None, "layer", "model")),
+        "base.decoder.layers.mlp.wi_1.kernel": ("model.layers.*.mlp.up_proj.weight", (None, "layer", "model")),
+        "base.decoder.layers.mlp.wo.kernel": ("model.layers.*.mlp.down_proj.weight", ("model", "layer", None)),
+        # MOE LAYERS (Map to INTERNAL keys to bypass loader stacking)
+        # Since MaxText experts are already fused/stacked, we map directly to the
+        # internal `tpu-inference` param names. The loader will fail to find
+        # "experts.{i}" in the name and fall back to loading these directly,
+        # which is exactly what we want for performance.
+        # Shared Experts
+        "base.decoder.layers.DeepSeekMoeBlock_0.shared_experts.wi_0.kernel": (
+            "layers.*.shared_experts.kernel_gating_DF",
+            (None, "layer", "model"),
+        ),
+        "base.decoder.layers.DeepSeekMoeBlock_0.shared_experts.wi_1.kernel": (
+            "layers.*.shared_experts.kernel_up_proj_DF",
+            (None, "layer", "model"),
+        ),
+        "base.decoder.layers.DeepSeekMoeBlock_0.shared_experts.wo.kernel": (
+            "layers.*.shared_experts.kernel_down_proj_FD",
+            ("model", "layer", None),
+        ),
+        # Router
+        "base.decoder.layers.DeepSeekMoeBlock_0.MoeBlock_0.gate.kernel": (
+            "layers.*.custom_module.router.kernel_DE",
+            (None, "layer", "model"),
+        ),
+        "base.decoder.layers.DeepSeekMoeBlock_0.MoeBlock_0.gate.bias": (
+            "layers.*.custom_module.router.bias_E",
+            (None, "layer", "model"),
+        ),
+        # Routed Experts (Fused)
+        "base.decoder.layers.DeepSeekMoeBlock_0.MoeBlock_0.wi_0": (
+            "layers.*.custom_module.kernel_gating_EDF",
+            ("expert", "layer", None, "model"),
+        ),
+        "base.decoder.layers.DeepSeekMoeBlock_0.MoeBlock_0.wi_1": (
+            "layers.*.custom_module.kernel_up_proj_EDF",
+            ("expert", "layer", None, "model"),
+        ),
+        "base.decoder.layers.DeepSeekMoeBlock_0.MoeBlock_0.wo": (
+            "layers.*.custom_module.kernel_down_proj_EFD",
+            ("expert", "layer", "model", None),
+        ),
+        # MTP BLOCK (Included for completeness, but typically skipped by current loader)
+        "base.mtp_block.mtp_layer_1.embedding_norm.scale": ("mtp_block.layer.pre_norm.scale", (None,)),
+        "base.mtp_block.mtp_layer_1.hidden_state_norm.scale": ("mtp_block.layer.post_norm.scale", (None,)),
+        "base.mtp_block.mtp_layer_1.projection_layer.kernel": ("mtp_block.layer.projection.kernel", (None, "model")),
+    }
+    return mapping
diff --git a/src/MaxText/integration/tunix/weight_mapping/gpt_oss.py b/src/MaxText/integration/tunix/weight_mapping/gpt_oss.py
@@ -0,0 +1,157 @@
+# Copyright 2023–2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Mapping MaxText GPT-OSS (MoE) weights to vLLM/tpu-inference keys."""
+
+from dataclasses import dataclass
+import logging
+from typing import Dict, Optional, Tuple
+import jax
+
+
+@dataclass
+class GPT_OSS_VLLM_MAPPING:
+  """
+  Mapping definition from MaxText GPT-OSS (Scanned/Interleaved) to vLLM JAX NNX.
+  Supports:
+  - Modulo Interleaving (e.g., Block 0 -> Layers 0, 2, 4...)
+  """
+
+  @staticmethod
+  def lora_to_hf_mappings():
+    """Provides the mapping for LoRA (Low-Rank Adaptation) weights.
+    Returns:
+        None, as LoRA mappings are not defined for this model.
+    """
+    return None
+
+  @staticmethod
+  def to_hf_hook_fns():
+    """Returns hook functions to fuse interleaved weights."""
+    return {}
+
+  @staticmethod
+  def to_hf_transpose_keys():
+    """Returns keys that need to be transposed."""
+    return {}
+
+  @staticmethod
+  def to_hf_mapping(
+      layer_cycle_interval: int = 2, total_num_layers: int = 36, interleave_style: str = "modulo"
+  ) -> Dict[str, Tuple[str, Tuple[Optional[str], ...]]]:
+    """Returns the weight mapping for the model.
+    Args:
+        layer_cycle_interval: The interval at which layers are cycled.
+        total_num_layers: The total number of layers in the model.
+        interleave_style: The style of interleaving used for the layers.
+    Returns:
+        A dictionary mapping MaxText parameter names to vLLM parameter names.
+    """
+
+    mapping = {}
+
+    # --- 1. Global Parameters ---
+    mapping.update(
+        {
+            "base.token_embedder.embedding": ("embedder.input_embedding_table_VD", ("model", None)),
+            "base.decoder.decoder_norm.scale": ("final_norm.scale", (None,)),
+            "base.decoder.logits_dense.kernel": ("lm_head.input_embedding_table_DV", (None, "model")),
+        }
+    )
+
+    # --- 2. Layer Mapping Loop ---
+    layers_per_block = total_num_layers // layer_cycle_interval
+
+    for block_idx in range(layer_cycle_interval):
+      src_block = f"base.decoder.layers.layers_{block_idx}"
+      if interleave_style == "modulo":
+        target_indices = range(block_idx, total_num_layers, layer_cycle_interval)
+      else:
+        start = block_idx * layers_per_block
+        target_indices = range(start, start + layers_per_block)
+
+      regex_indices = "|".join(map(str, target_indices))
+      layer_regex = f"layers\.({regex_indices})"
+
+      # --- 3. Block Mappings (Standard) ---
+      mapping.update(
+          {
+              f"{src_block}.pre_self_attention_layer_norm.scale": (
+                  f"{layer_regex}.pre_attention_norm.scale",
+                  (None, "layer"),
+              ),
+              f"{src_block}.post_self_attention_layer_norm.scale": (f"{layer_regex}.pre_mlp_norm.scale", (None, "layer")),
+              f"{src_block}.GptOssAttention.query.kernel": (
+                  f"{layer_regex}.attn.kernel_q_DNH",
+                  (None, "layer", "model", None),
+              ),
+              f"{src_block}.GptOssAttention.key.kernel": (
+                  f"{layer_regex}.attn.kernel_k_DKH",
+                  (None, "layer", "model", None),
+              ),
+              f"{src_block}.GptOssAttention.value.kernel": (
+                  f"{layer_regex}.attn.kernel_v_DKH",
+                  (None, "layer", "model", None),
+              ),
+              f"{src_block}.GptOssAttention.out.kernel": (
+                  f"{layer_regex}.attn.kernel_o_proj_NHD",
+                  ("model", "layer", None, None),
+              ),
+              f"{src_block}.GptOssAttention.query.bias": (f"{layer_regex}.attn.bias_q_NH", (None, "layer", None)),
+              f"{src_block}.GptOssAttention.key.bias": (f"{layer_regex}.attn.bias_k_KH", (None, "layer", None)),
+              f"{src_block}.GptOssAttention.value.bias": (f"{layer_regex}.attn.bias_v_KH", (None, "layer", None)),
+              f"{src_block}.GptOssAttention.out.bias": (f"{layer_regex}.attn.bias_o_D", (None, "layer")),
+              f"{src_block}.GptOssAttention.sinks": (f"{layer_regex}.attn.sinks_N", (None, "layer")),
+          }
+      )
+
+      # MoE Router
+      mapping.update(
+          {
+              f"{src_block}.GptOssMlp.gate.kernel": (
+                  f"{layer_regex}.custom_module.router.kernel_DE",
+                  (None, "layer", "model"),
+              ),
+              f"{src_block}.GptOssMlp.gate.bias": (f"{layer_regex}.custom_module.router.bias_E", ("model", "layer")),
+          }
+      )
+
+      # --- MOE EXPERTS ---
+      # Separate gate_proj (wi_0) and up_proj (wi_1) kernels and biases.
+
+      # MLP Gate Projection (wi_0)
+      mapping.update(
+          {
+              f"{src_block}.GptOssMlp.wi_0": (f"{layer_regex}.custom_module.gate_proj_kernel", ("model", "layer", None)),
+              f"{src_block}.GptOssMlp.wi_0_bias": (f"{layer_regex}.custom_module.gate_proj_bias", ("model", "layer")),
+          }
+      )
+
+      # MLP Up Projection (wi_1)
+      mapping.update(
+          {
+              f"{src_block}.GptOssMlp.wi_1": (f"{layer_regex}.custom_module.up_proj_kernel", ("model", "layer", None)),
+              f"{src_block}.GptOssMlp.wi_1_bias": (f"{layer_regex}.custom_module.up_proj_bias", ("model", "layer")),
+          }
+      )
+
+      # MLP Down Projection (wo)
+      mapping.update(
+          {
+              f"{src_block}.GptOssMlp.wo": (f"{layer_regex}.custom_module.mlp2_weight_EFD", ("model", "layer", None)),
+              f"{src_block}.GptOssMlp.wo_bias": (f"{layer_regex}.custom_module.mlp2_bias_ED", ("model", "layer")),
+          }
+      )
+
+    return mapping
