tpu-inference 0.0.1rc1__py3-none-any.whl

tpu_inference/models/jax/llama_eagle3.py

@@ -0,0 +1,336 @@
+from typing import List, Tuple
+
+import jax
+import jax.numpy as jnp
+from flax import nnx
+from jax.sharding import Mesh
+from transformers import LlamaConfig
+from vllm.config import VllmConfig
+
+from tpu_inference.layers.common.attention_metadata import AttentionMetadata
+from tpu_inference.logger import init_logger
+from tpu_inference.models.jax.llama3 import LlamaDecoderLayer
+from tpu_inference.models.jax.utils.weight_utils import (MetadataMap,
+                                                         get_default_maps,
+                                                         load_hf_weights)
+
+logger = init_logger(__name__)
+
+init_fn = nnx.initializers.uniform()
+
+
+class Eagle3LlamaDecoderLayer(LlamaDecoderLayer):
+
+    def __init__(self, config: LlamaConfig, dtype: jnp.dtype, rng: nnx.Rngs,
+                 mesh: Mesh, kv_cache_dtype: str):
+        super().__init__(config,
+                         dtype=dtype,
+                         rng=rng,
+                         mesh=mesh,
+                         kv_cache_dtype=kv_cache_dtype)
+        self.config = config
+        # Override qkv
+        hidden_size = 2 * self.self_attn.hidden_size
+        self.self_attn.q_proj = nnx.Einsum(
+            "TD,DNH->TNH",
+            (hidden_size, self.self_attn.num_heads, self.self_attn.head_dim),
+            param_dtype=dtype,
+            dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model", None)),
+            rngs=rng,
+        )
+        self.self_attn.k_proj = nnx.Einsum(
+            "TD,DKH->TKH",
+            (hidden_size, self.self_attn.num_kv_heads,
+             self.self_attn.head_dim),
+            param_dtype=dtype,
+            dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model", None)),
+            rngs=rng,
+        )
+        self.self_attn.v_proj = nnx.Einsum(
+            "TD,DKH->TKH",
+            (hidden_size, self.self_attn.num_kv_heads,
+             self.self_attn.head_dim),
+            param_dtype=dtype,
+            dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model", None)),
+            rngs=rng,
+        )
+        # Override input layernorm and specify dtype to avoid unexpected upcasting.
+        self.input_layernorm = nnx.RMSNorm(
+            config.hidden_size,
+            epsilon=config.rms_norm_eps,
+            param_dtype=dtype,
+            dtype=dtype,
+            scale_init=nnx.with_partitioning(init_fn, (None, )),
+            rngs=rng,
+        )
+        self.hidden_norm = nnx.RMSNorm(
+            config.hidden_size,
+            epsilon=config.rms_norm_eps,
+            param_dtype=dtype,
+            scale_init=nnx.with_partitioning(init_fn, (None, )),
+            rngs=rng,
+        )
+
+    def _norm_before_residual(
+            self, hidden_states: jax.Array) -> tuple[jax.Array, jax.Array]:
+        hidden_states = self.hidden_norm(hidden_states)
+        residual = hidden_states
+        return hidden_states, residual
+
+    def _norm_after_residual(
+            self, hidden_states: jax.Array) -> tuple[jax.Array, jax.Array]:
+        residual = hidden_states
+        hidden_states = self.hidden_norm(hidden_states)
+        return hidden_states, residual
+
+    def __call__(
+        self,
+        kv_cache: jax.Array,
+        embeds: jax.Array,
+        hidden_states: jax.Array,
+        attention_metadata: AttentionMetadata,
+    ) -> Tuple[jax.Array, jax.Array, jax.Array]:
+        embeds = self.input_layernorm(embeds)
+        if getattr(self.config, "norm_before_residual", False):
+            hidden_states, residual = self._norm_before_residual(
+                hidden_states=hidden_states)
+        else:
+            hidden_states, residual = self._norm_after_residual(
+                hidden_states=hidden_states)
+        hidden_states = jnp.concatenate([embeds, hidden_states], axis=-1)
+
+        kv_cache, attn_output = self.self_attn(
+            kv_cache,
+            hidden_states,
+            attention_metadata,
+        )
+
+        # TODO(ranlihao): Check if this residual connection is correct.
+        hidden_states = attn_output + residual
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        mlp_output = self.mlp(hidden_states)
+
+        return kv_cache, mlp_output, residual
+
+
+class Eagle3LlamaModel(nnx.Module):
+
+    def __init__(self, vllm_config: VllmConfig, rng: nnx.Rngs, mesh: Mesh):
+        super().__init__()
+        hf_config = vllm_config.speculative_config.draft_model_config.hf_config
+        dtype: jnp.dtype = jnp.bfloat16
+
+        self.embed_tokens = nnx.Embed(
+            num_embeddings=hf_config.vocab_size,
+            features=hf_config.hidden_size,
+            param_dtype=dtype,
+            embedding_init=nnx.with_partitioning(init_fn, ("model", None)),
+            rngs=rng,
+        )
+
+        self.layers = [
+            Eagle3LlamaDecoderLayer(
+                config=hf_config,
+                dtype=dtype,
+                rng=rng,
+                mesh=mesh,
+                # TODO (jacobplatin): we should refactor this to pass a dtype (or config) directly
+                kv_cache_dtype=vllm_config.cache_config.cache_dtype)
+        ]
+
+        if hasattr(hf_config, "target_hidden_size"):
+            input_size = hf_config.target_hidden_size * 3
+        else:
+            input_size = hf_config.hidden_size * 3
+
+        self.fc = nnx.Linear(
+            in_features=input_size,
+            out_features=hf_config.hidden_size,
+            use_bias=False,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model")),
+            rngs=rng,
+        )
+
+        self.norm = nnx.RMSNorm(
+            hf_config.hidden_size,
+            epsilon=hf_config.rms_norm_eps,
+            param_dtype=dtype,
+            scale_init=nnx.with_partitioning(init_fn, (None, )),
+            rngs=rng,
+        )
+
+    def __call__(
+        self,
+        kv_caches: List[jax.Array],
+        input_ids: jax.Array,
+        hidden_states: jax.Array,
+        attention_metadata: AttentionMetadata,
+    ) -> Tuple[List[jax.Array], jax.Array, List[jax.Array]]:
+        embeds = self.embed_tokens(input_ids)
+        assert hidden_states.shape[-1] == embeds.shape[-1]
+
+        assert len(self.layers) == 1
+        # The first N - 1 KV caches are for the target model, and the last one is for the draft model.
+        # N is the number of layers in the target model.
+        # The draft model has only 1 layer.
+        kv_caches[-1], hidden_states, residual = self.layers[0](
+            kv_caches[-1],
+            embeds,
+            hidden_states,
+            attention_metadata,
+        )
+
+        # TODO(ranlihao): Check if this residual connection is correct.
+        hidden_states = hidden_states + residual
+        residual = hidden_states
+        hidden_states = self.norm(hidden_states)
+        return kv_caches, hidden_states, [residual]
+
+
+def update_reshape_map_for_eagle3(vllm_config: VllmConfig,
+                                  metadata_map: MetadataMap):
+    model_config = vllm_config.speculative_config.draft_model_config
+    hf_config = model_config.hf_config
+
+    num_heads = hf_config.num_attention_heads
+    num_kv_heads = hf_config.num_key_value_heads
+    hidden_size = hf_config.hidden_size
+    head_dim_original = model_config.get_head_size()
+
+    metadata_map.reshape_map.update({
+        "q_proj": (num_heads, head_dim_original, 2 * hidden_size),
+        "k_proj": (num_kv_heads, head_dim_original, 2 * hidden_size),
+        "v_proj": (num_kv_heads, head_dim_original, 2 * hidden_size),\
+    })
+
+
+class EagleLlama3ForCausalLM(nnx.Module):
+
+    def __init__(self, vllm_config: VllmConfig, rng_key: jax.Array,
+                 mesh: Mesh):
+        nnx.Module.__init__(self)
+        self.vllm_config = vllm_config
+        self.rng = nnx.Rngs(rng_key)
+        self.mesh = mesh
+        dtype: jnp.dtype = jnp.bfloat16
+
+        spec_config = vllm_config.speculative_config
+        assert spec_config is not None
+        model_config = spec_config.draft_model_config
+        assert model_config is not None
+        hf_config = model_config.hf_config
+
+        self.model = Eagle3LlamaModel(
+            vllm_config=vllm_config,
+            rng=self.rng,
+            mesh=mesh,
+        )
+
+        self.lm_head = nnx.Linear(
+            hf_config.hidden_size,
+            hf_config.draft_vocab_size,
+            use_bias=False,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model")),
+            rngs=self.rng,
+        )
+
+        self.draft_id_to_target_id = nnx.Param(jnp.zeros(
+            hf_config.draft_vocab_size, dtype=jnp.int32),
+                                               sharding=(None, ))
+
+    def __call__(
+        self,
+        kv_caches: List[jax.Array],
+        input_ids: jax.Array,
+        hidden_states: jax.Array,
+        attention_metadata: AttentionMetadata,
+    ) -> Tuple[List[jax.Array], jax.Array, List[jax.Array]]:
+        return self.model(
+            kv_caches,
+            input_ids,
+            hidden_states,
+            attention_metadata,
+        )
+
+    def compute_logits(self, hidden_states: jax.Array) -> jax.Array:
+        logits = self.lm_head(hidden_states)
+
+        target_vocab_size = self.vllm_config.model_config.get_vocab_size()
+        draft_vocab_size = self.vllm_config.speculative_config.draft_model_config.hf_config.draft_vocab_size
+
+        base = jnp.arange(draft_vocab_size, dtype=jnp.int32)
+        targets = base + self.draft_id_to_target_id.value
+
+        logits_new = jnp.full((logits.shape[0], target_vocab_size),
+                              -jnp.inf,
+                              dtype=logits.dtype)
+
+        logits_new = logits_new.at[:, targets].set(logits)
+
+        return logits_new
+
+    def combine_hidden_states(self, hidden_states: jax.Array) -> jax.Array:
+        return self.model.fc(hidden_states)
+
+    def load_weights(self, rng_key: jax.Array):
+        # Create a new Rngs object for the draft model to avoid sharing RNG state
+        self.rng = jax.random.key(self.vllm_config.model_config.seed)
+        spec_config = self.vllm_config.speculative_config
+        assert spec_config is not None
+
+        mappings = {
+            "midlayer.input_layernorm": "model.layers.0.input_layernorm.scale",
+            "midlayer.hidden_norm": "model.layers.0.hidden_norm.scale",
+            "midlayer.mlp.down_proj": "model.layers.0.mlp.down_proj.kernel",
+            "midlayer.mlp.gate_proj": "model.layers.0.mlp.gate_proj.kernel",
+            "midlayer.mlp.up_proj": "model.layers.0.mlp.up_proj.kernel",
+            "midlayer.post_attention_layernorm":
+            "model.layers.0.post_attention_layernorm.scale",
+            "midlayer.self_attn.k_proj":
+            "model.layers.0.self_attn.k_proj.kernel",
+            "midlayer.self_attn.o_proj":
+            "model.layers.0.self_attn.o_proj.kernel",
+            "midlayer.self_attn.q_proj":
+            "model.layers.0.self_attn.q_proj.kernel",
+            "midlayer.self_attn.v_proj":
+            "model.layers.0.self_attn.v_proj.kernel",
+            "norm": "model.norm.scale",
+            "fc": "model.fc.kernel",
+            "lm_head": "lm_head.kernel",
+            "d2t": "draft_id_to_target_id",
+            "embed_tokens":
+            "model.embed_tokens.embedding",  # Some checkpoints need this
+        }
+
+        # Define keys to keep in original dtype (e.g., float32 for stability)
+        keep_original_dtype_keys_regex = [
+            r".*d2t.*",
+        ]
+
+        metadata_map = get_default_maps(
+            self.vllm_config.speculative_config.draft_model_config, self.mesh,
+            mappings)
+
+        update_reshape_map_for_eagle3(self.vllm_config, metadata_map)
+
+        load_hf_weights(
+            vllm_config=self.vllm_config,
+            model=self,
+            metadata_map=metadata_map,
+            mesh=self.mesh,
+            is_draft_model=True,
+            keep_original_dtype_keys_regex=keep_original_dtype_keys_regex)
+
+        # If the embedding is not initialized, initialize it with a dummy array here to pass jit compilation. The real weights will be shared from the target model in eagle3 class.
+        if isinstance(self.model.embed_tokens.embedding.value,
+                      jax.ShapeDtypeStruct):
+            self.model.embed_tokens.embedding.value = jnp.zeros(
+                self.model.embed_tokens.embedding.shape,
+                dtype=self.model.embed_tokens.embedding.dtype,
+            )

tpu_inference/models/jax/llama_guard_4.py

@@ -0,0 +1,361 @@
+import re
+from typing import Any, List, Optional, Tuple
+
+import jax
+import jax.numpy as jnp
+import torch
+from flax import nnx
+from flax.typing import PRNGKey
+from jax.sharding import Mesh
+from jax.sharding import PartitionSpec as P
+from vllm.config import VllmConfig
+
+from tpu_inference.layers.jax.attention.attention import AttentionMetadata
+from tpu_inference.layers.jax.attention.llama4_attention import Llama4Attention
+from tpu_inference.layers.jax.constants import KVCacheType
+from tpu_inference.layers.jax.layers import DenseFFW, Embedder, LMhead, RMSNorm
+from tpu_inference.layers.jax.misc import shard_put
+from tpu_inference.layers.jax.transformer_block import TransformerBlock
+from tpu_inference.logger import init_logger
+from tpu_inference.models.jax.utils.weight_utils import (
+    get_param, model_weights_generator, print_param_info, reshape_params,
+    transpose_params)
+
+logger = init_logger(__name__)
+
+
+class LlamaGuard4ForCausalLM(nnx.Module):
+
+    def __init__(self,
+                 vllm_config: VllmConfig,
+                 rng: PRNGKey,
+                 mesh: Mesh,
+                 force_random_weights: bool = False):
+        logger.warning(
+            "🚨🚨🚨WARNING🚨🚨🚨 🚨🚨🚨WARNING🚨🚨🚨 🚨🚨🚨WARNING🚨🚨🚨\n"
+            "Llama Guard 4 (JAX) is WIP: Only the text modality is currently implemented.  "
+            "Multimodal inputs will fail.\n"
+            "🚨🚨🚨WARNING🚨🚨🚨 🚨🚨🚨WARNING🚨🚨🚨 🚨🚨🚨WARNING🚨🚨🚨")
+        assert mesh is not None
+
+        self.vllm_config = vllm_config
+        self.vllm_config.model_config.dtype = torch.bfloat16
+        model_config = vllm_config.model_config
+        text_config = model_config.hf_config.text_config
+
+        self.mesh = mesh
+        self.is_verbose = getattr(self.vllm_config.additional_config,
+                                  "is_verbose", False)
+
+        self.use_qk_norm = getattr(text_config, "use_qk_norm", True)
+
+        vocab_size = model_config.get_vocab_size()
+        self.hidden_size = model_config.get_hidden_size()
+
+        self.dtype: jnp.dtype = jnp.bfloat16
+
+        self.num_layers: int = getattr(text_config, "num_layers", 48)
+        hidden_act: str = getattr(text_config, "hidden_act", "silu")
+
+        rms_norm_eps = getattr(text_config, "rms_norm_eps", 1e-5)
+        self.num_attention_heads = getattr(text_config, "num_attention_heads",
+                                           40)
+        self.num_key_value_heads = getattr(text_config, "num_key_value_heads",
+                                           8)
+        self.head_dim = getattr(text_config, "head_dim", 128)
+
+        intermediate_size = getattr(text_config, "intermediate_size", 8192)
+
+        self.rope_theta_text = getattr(text_config, "rope_theta", 500000.0)
+        self.rope_scaling = getattr(text_config, "rope_scaling")
+
+        self.rng = nnx.Rngs(rng)
+
+        self.embedder = Embedder(
+            vocab_size=vocab_size,
+            hidden_size=self.hidden_size,
+            dtype=self.dtype,
+            vd_sharding=(('data', 'model'), None),
+            rngs=self.rng,
+            random_init=force_random_weights,
+        )
+
+        self.layers = []
+
+        for i in range(self.num_layers):
+            use_attention_rope = True
+
+            custom_module = DenseFFW(dtype=self.dtype,
+                                     hidden_act=hidden_act,
+                                     hidden_size=self.hidden_size,
+                                     intermediate_size=intermediate_size,
+                                     random_init=force_random_weights,
+                                     rngs=self.rng,
+                                     df_sharding=P(None, 'model'),
+                                     fd_sharding=P('model', None),
+                                     activation_ffw_td=P('data', None))
+
+            attn = Llama4Attention(
+                hidden_size=self.hidden_size,
+                dtype=self.dtype,
+                num_attention_heads=self.num_attention_heads,
+                num_key_value_heads=self.num_key_value_heads,
+                head_dim=self.head_dim,
+                rope_theta=self.rope_theta_text,
+                rope_scaling={
+                    "scale_factor":
+                    self.rope_scaling["factor"],
+                    "low_freq_factor":
+                    self.rope_scaling["low_freq_factor"],
+                    "high_freq_factor":
+                    self.rope_scaling["high_freq_factor"],
+                    "original_max_position_embeddings":
+                    self.rope_scaling["original_max_position_embeddings"]
+                },
+                rngs=self.rng,
+                rope_input_ordering="interleaved",
+                # TODO (jacobplatin): we should refactor this to pass a dtype (or config) directly
+                kv_cache_dtype=vllm_config.cache_config.cache_dtype,
+                temperature_tuning=True,
+                temperature_tuning_scale=0.1,
+                temperature_tuning_floor_scale=8192,
+                use_qk_norm=self.use_qk_norm,
+                attention_chunk_size=None if use_attention_rope else 8192,
+                mesh=self.mesh,
+                random_init=force_random_weights,
+                activation_attention_td=('data', 'model'),
+                activation_q_td=('data', 'model'),
+                query_tnh=P('data', 'model', None),
+                keyvalue_skh=P('data', 'model', None),
+                activation_attention_out_td=('data', 'model'),
+                attn_o_tnh=P('data', 'model', None),
+                dnh_sharding=(None, 'model', None),
+                dkh_sharding=(None, 'model', None),
+                nhd_sharding=('model', None, None),
+            )
+
+            pre_attention_norm = RMSNorm(
+                dims=self.hidden_size,
+                random_init=force_random_weights,
+                epsilon=rms_norm_eps,
+                rngs=self.rng,
+                activation_ffw_td=('data', None),
+                with_scale=True,
+                dtype=self.dtype,
+            )
+
+            pre_mlp_norm = RMSNorm(
+                dims=self.hidden_size,
+                activation_ffw_td=('data', None),
+                epsilon=rms_norm_eps,
+                rngs=self.rng,
+                with_scale=True,
+                dtype=self.dtype,
+                random_init=force_random_weights,
+            )
+
+            block = TransformerBlock(custom_module=custom_module,
+                                     attn=attn,
+                                     pre_attention_norm=pre_attention_norm,
+                                     pre_mlp_norm=pre_mlp_norm,
+                                     use_attention_rope=use_attention_rope)
+            self.layers.append(block)
+
+        self.final_norm = RMSNorm(
+            dims=self.hidden_size,
+            activation_ffw_td=P(),
+            epsilon=rms_norm_eps,
+            rngs=self.rng,
+            with_scale=True,
+            dtype=self.dtype,
+            random_init=force_random_weights,
+        )
+
+        self.lm_head = LMhead(vocab_size=vocab_size,
+                              hidden_size=self.hidden_size,
+                              dtype=self.dtype,
+                              rngs=self.rng,
+                              vd_sharding=(('data', 'model'), None),
+                              dv_sharding=(None, ('data', 'model')),
+                              random_init=force_random_weights)
+        if self.is_verbose:
+            self._print_model_architecture()
+
+    def _print_model_architecture(self):
+
+        logger.info("### Embedding ###")
+        nnx.display(self.embedder)
+
+        logger.info("\n### Layers ###")
+        for i, layer in enumerate(self.layers):
+            logger.info(f"\n--- Layer {i} ---")
+            nnx.display(layer)
+
+        logger.info("\n### LM Head ###")
+        nnx.display(self.lm_head)
+
+    def load_weights(self, rng: jax.Array, cache_dir: Optional[str] = None):
+        self.rng = nnx.Rngs(rng)
+
+        weight_loader = LlamaGuard4WeightLoader(
+            vllm_config=self.vllm_config,
+            hidden_size=self.hidden_size,
+            attn_heads=self.num_attention_heads,
+            num_key_value_heads=self.num_key_value_heads,
+            attn_head_dim=self.head_dim)
+        weight_loader.load_weights(self)
+
+    def __call__(
+        self,
+        kv_caches: List[jax.Array],
+        input_ids: jax.Array,
+        attention_metadata: AttentionMetadata,
+        inputs_embeds: Optional[jax.Array] = None,
+        layer_metadata_tuple: Optional[Tuple] = None,
+        lora_metadata: Optional[Any] = None,
+        *args,
+    ) -> Tuple[List[KVCacheType], jax.Array]:
+        is_prefill = False
+
+        if inputs_embeds is not None:
+            x_TD = inputs_embeds
+        elif input_ids is not None:
+            x_TD = self.embedder.encode(input_ids)
+        else:
+            raise ValueError(
+                "Cannot run forward pass: Both input_ids and inputs_embeds are None."
+            )
+
+        for (i, block) in enumerate(self.layers):
+            kv_cache = kv_caches[i]
+            new_kv_cache, x_TD = block(x_TD, is_prefill, kv_cache,
+                                       attention_metadata)
+            jax.block_until_ready(x_TD)
+            kv_caches[i] = new_kv_cache
+
+        final_activation_TD = self.final_norm(x_TD)
+
+        return kv_caches, final_activation_TD, []
+
+    def compute_logits(self, hidden_states: jax.Array) -> jax.Array:
+        logits_TV = jnp.dot(hidden_states,
+                            self.lm_head.input_embedding_table_DV.value)
+        return logits_TV
+
+    def get_input_embeddings(
+            self,
+            input_ids: jax.Array,
+            multimodal_embeddings: Optional[List[jax.Array]] = None
+    ) -> jax.Array:
+        """
+        Computes the embeddings for text input (used for input to fusion).
+        """
+        return self.embedder.encode(input_ids)
+
+
+class LlamaGuard4WeightLoader:
+
+    def __init__(self, vllm_config: VllmConfig, hidden_size, attn_heads,
+                 num_key_value_heads, attn_head_dim):
+        self.names_and_weights_generator = model_weights_generator(
+            model_name_or_path=vllm_config.model_config.model,
+            framework="flax",
+            filter_regex="language_model",
+            download_dir=vllm_config.load_config.download_dir)
+        self.is_verbose = getattr(vllm_config.additional_config, "is_verbose",
+                                  False)
+        self._transpose_map = {
+            "q_proj": (2, 0, 1),
+            "k_proj": (2, 0, 1),
+            "v_proj": (2, 0, 1),
+            "o_proj": (1, 2, 0),
+            "lm_head": (1, 0),
+            "feed_forward.down_proj": (1, 0),
+            "feed_forward.gate_proj": (1, 0),
+            "feed_forward.up_proj": (1, 0),
+            "mlp.down_proj": (1, 0),
+            "mlp.gate_proj": (1, 0),
+            "mlp.up_proj": (1, 0),
+        }
+        self._weight_shape_map = {
+            "q_proj": (attn_heads, attn_head_dim, hidden_size),
+            "k_proj": (num_key_value_heads, attn_head_dim, hidden_size),
+            "v_proj": (num_key_value_heads, attn_head_dim, hidden_size),
+            "o_proj": (hidden_size, attn_heads, attn_head_dim),
+        }
+
+        self._loaded_to_standardized_keys = {
+            "language_model.model.embed_tokens.weight":
+            "embedder.input_embedding_table_VD",
+            "language_model.lm_head.weight":
+            "lm_head.input_embedding_table_DV",
+            "language_model.model.norm.weight":
+            "final_norm.scale",
+            "language_model.model.layers.*.input_layernorm.weight":
+            "layers.*.pre_attention_norm.scale",
+            "language_model.model.layers.*.post_attention_layernorm.weight":
+            "layers.*.pre_mlp_norm.scale",
+            "language_model.model.layers.*.self_attn.q_proj.weight":
+            "layers.*.attn.kernel_q_proj_DNH",
+            "language_model.model.layers.*.self_attn.k_proj.weight":
+            "layers.*.attn.kernel_k_proj_DKH",
+            "language_model.model.layers.*.self_attn.v_proj.weight":
+            "layers.*.attn.kernel_v_proj_DKH",
+            "language_model.model.layers.*.self_attn.o_proj.weight":
+            "layers.*.attn.kernel_o_proj_NHD",
+            "language_model.model.layers.*.feed_forward.gate_proj.weight":
+            "layers.*.custom_module.kernel_gating_DF",
+            "language_model.model.layers.*.feed_forward.up_proj.weight":
+            "layers.*.custom_module.kernel_up_proj_DF",
+            "language_model.model.layers.*.feed_forward.down_proj.weight":
+            "layers.*.custom_module.kernel_down_proj_FD",
+        }
+
+    def map_loaded_to_standardized_name(self, loaded_key: str) -> str:
+        if "layer" in loaded_key:
+            layer_num = re.search(r"layers\.(\d+)", loaded_key).group(1)
+            layer_key = re.sub(r"layers\.\d+", "layers.*", loaded_key)
+            mapped_key = self._loaded_to_standardized_keys.get(
+                layer_key, loaded_key)
+            mapped_key = re.sub(r"layers\.\*", f"layers.{layer_num}",
+                                mapped_key)
+        else:
+            mapped_key = self._loaded_to_standardized_keys.get(
+                loaded_key, loaded_key)
+        return mapped_key
+
+    def load_weights(self, model_for_loading: nnx.Module):
+        model_params = nnx.state(model_for_loading)
+        with jax.default_device(jax.devices("cpu")[0]):
+            for loaded_name, loaded_weight in self.names_and_weights_generator:
+                if loaded_name.endswith(".bias"):
+                    continue
+                if "vision_model" in loaded_name or "multi_modal_projector" in loaded_name:
+                    continue
+
+                mapped_name = self.map_loaded_to_standardized_name(loaded_name)
+                model_weight = get_param(model_params, mapped_name)
+
+                if not loaded_name.endswith(".bias"):
+                    # For other layers, continue to use the transpose_params helper.
+                    loaded_weight = reshape_params(loaded_name, loaded_weight,
+                                                   self._weight_shape_map)
+                    loaded_weight = transpose_params(loaded_name,
+                                                     loaded_weight,
+                                                     self._transpose_map)
+                if model_weight.value.shape != loaded_weight.shape:
+                    raise ValueError(
+                        f"Loaded shape for {loaded_name}: {loaded_weight.shape} "
+                        f"does not match model shape for {mapped_name}: {model_weight.value.shape}!"
+                    )
+                logger.debug(
+                    f"Transformed parameter {loaded_name} to {mapped_name}: {loaded_weight.shape} --> {model_weight.value.shape}"
+                )
+
+                model_weight.value = shard_put(loaded_weight,
+                                               model_weight.sharding,
+                                               mesh=model_for_loading.mesh)
+                if self.is_verbose:
+                    print_param_info(model_weight, loaded_name)
+
+        nnx.update(model_for_loading, model_params)