tpu-inference 0.11.1__py3-none-any.whl

tpu_inference/models/jax/llama4.py

@@ -0,0 +1,473 @@
+import re
+from typing import List, Optional, Tuple
+
+import jax
+import jax.numpy as jnp
+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.moe.moe import MoE, Router
+from tpu_inference.layers.jax.transformer_block import \
+    SharedExpertsTransformerBlock
+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 Llama4ForCausalLM(nnx.Module):
+
+    def __init__(self,
+                 vllm_config: VllmConfig,
+                 rng: PRNGKey,
+                 mesh: Mesh,
+                 force_random_weights: bool = False):
+        assert mesh is not None
+
+        self.vllm_config = vllm_config
+        model_config = vllm_config.model_config
+        text_config = model_config.hf_config.text_config
+
+        self.rng = nnx.Rngs(rng)
+        self.mesh = mesh
+        self.is_verbose = getattr(self.vllm_config.additional_config,
+                                  "is_verbose", False)
+
+        # Currently the runner will always set a mesh, so the custom default sharding (when
+        #  no sharding is set in vllm config) doesn't take effect.
+        # TODO(fhzhang): figure out whether we need to actually enable this.
+        #    strategy_dict = {"tensor_parallelism": 4, "expert_parallelism": 2}
+
+        # TODO(fhzhang): remove these once we confirm that the values we get from config are good.
+        # self.hidden_size: int = 5120
+        # vocab_size = 202048
+        self.vocab_size = model_config.get_vocab_size()
+        self.hidden_size = model_config.get_hidden_size()
+
+        dtype: jnp.dtype = jnp.bfloat16
+
+        self.num_layers: int = getattr(text_config, "num_hidden_layers", 48)
+
+        self.intermediate_size_moe: int = getattr(text_config,
+                                                  "intermediate_size", 8192)
+        self.intermediate_size_mlp = getattr(text_config,
+                                             "intermediate_size_mlp", 16384)
+
+        # num_local_experts: uses 16 experts for Llama-4-Scout-17B-16E-Instruct and uses 128 experts Llama-4-Maverick-17B-128E-Instruct.
+        # The default value is set to 16 for compatibility with Llama-4-Scout.
+        self.num_local_experts: int = getattr(text_config, "num_local_experts",
+                                              16)
+        self.hidden_act: str = getattr(text_config, "hidden_act", "silu")
+        self.no_rope_layer_interval = getattr(text_config, "no_rope_layers",
+                                              [])
+
+        # interleave_moe_layer_step has a layer step of 2 to interleave MoE and dense layers for Llama-4-Maverick-17B-128E-Instruct.
+        # The default value is set to 1 for compatibility with Llama-4-Scout.
+        self.interleave_moe_layer_step = getattr(text_config,
+                                                 "interleave_moe_layer_step",
+                                                 1)
+
+        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)
+
+        self.num_shared_experts = getattr(text_config, "num_experts_per_tok",
+                                          1)
+        self.rms_norm_eps = getattr(text_config, "rms_norm_eps", 1e-5)
+
+        self.embedder = Embedder(vocab_size=self.vocab_size,
+                                 hidden_size=self.hidden_size,
+                                 dtype=dtype,
+                                 vd_sharding=(('data', 'expert', 'model'),
+                                              None),
+                                 rngs=self.rng,
+                                 random_init=force_random_weights)
+
+        self.layers = []
+
+        for i in range(self.num_layers):
+            # For Llama4-Scout, all layers are MoE layers.
+            # This can be adjusted for other variants.
+            is_moe_layer = (i + 1) % \
+                            self.interleave_moe_layer_step == 0
+
+            # Llama-4-Scout config: It has "no_rope_layers": []
+            use_attention_rope = (i + 1) not in self.no_rope_layer_interval
+
+            router = Router(dtype=dtype,
+                            hidden_size=self.hidden_size,
+                            num_experts=self.num_local_experts,
+                            num_experts_per_tok=1,
+                            router_act="sigmoid",
+                            rngs=self.rng,
+                            activation_ffw_td=('data', None),
+                            ed_sharding=(None, 'expert'),
+                            random_init=force_random_weights)
+
+            custom_module = MoE(
+                dtype=dtype,
+                num_local_experts=self.num_local_experts,
+                apply_expert_weight_before_computation=True,
+                hidden_size=self.hidden_size,
+                intermediate_size_moe=self.intermediate_size_moe,
+                hidden_act=self.hidden_act,
+                router=router,
+                rngs=self.rng,
+                activation_ffw_td=('data', None),
+                activation_ffw_ted=('data', 'expert', None),
+                edf_sharding=('expert', None, 'model'),
+                efd_sharding=('expert', 'model', None),
+                random_init=force_random_weights
+            ) if is_moe_layer else DenseFFW(
+                dtype=dtype,
+                hidden_act=self.hidden_act,
+                hidden_size=self.hidden_size,
+                intermediate_size=self.intermediate_size_mlp,
+                random_init=force_random_weights,
+                rngs=self.rng,
+                df_sharding=(None, 'model'),
+                fd_sharding=('model', None),
+                activation_ffw_td=('data', None))
+
+            attn = Llama4Attention(
+                hidden_size=self.hidden_size,
+                dtype=dtype,
+                kv_cache_dtype=vllm_config.cache_config.cache_dtype,
+                num_attention_heads=self.num_attention_heads,
+                num_key_value_heads=self.num_key_value_heads,
+                head_dim=self.head_dim,
+                rope_theta=500000.0,
+                # https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/blob/main/config.json
+                rope_scaling={
+                    "scale_factor": 16.0,
+                    "low_freq_factor": 1.0,
+                    "high_freq_factor": 1.0,
+                    "original_max_position_embeddings": 8192
+                },
+                rngs=self.rng,
+                rope_input_ordering="interleaved",
+                temperature_tuning=True,
+                temperature_tuning_scale=0.1,
+                temperature_tuning_floor_scale=8192,
+                use_qk_norm=True,
+                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),
+            )
+
+            shared_experts = DenseFFW(
+                dtype=dtype,
+                hidden_act=self.hidden_act,
+                hidden_size=self.hidden_size,
+                intermediate_size=self.num_shared_experts *
+                self.intermediate_size_moe,
+                rngs=self.rng,
+                random_init=force_random_weights,
+                df_sharding=(None, 'model'),
+                fd_sharding=('model', None),
+                activation_ffw_td=('data', None)) if is_moe_layer else None
+
+            pre_attention_norm = RMSNorm(
+                dims=self.hidden_size,
+                random_init=force_random_weights,
+                epsilon=self.rms_norm_eps,
+                rngs=self.rng,
+                with_scale=True,
+                dtype=dtype,
+            )
+
+            pre_mlp_norm = RMSNorm(
+                dims=self.hidden_size,
+                epsilon=self.rms_norm_eps,
+                rngs=self.rng,
+                with_scale=True,
+                dtype=dtype,
+                random_init=force_random_weights,
+            )
+
+            block = SharedExpertsTransformerBlock(
+                custom_module=custom_module,
+                attn=attn,
+                pre_attention_norm=pre_attention_norm,
+                pre_mlp_norm=pre_mlp_norm,
+                shared_experts=shared_experts,
+                use_attention_rope=use_attention_rope)
+            self.layers.append(block)
+
+        self.final_norm = RMSNorm(
+            dims=self.hidden_size,
+            epsilon=self.rms_norm_eps,
+            rngs=self.rng,
+            with_scale=True,
+            dtype=dtype,
+            random_init=force_random_weights,
+        )
+
+        self.lm_head = LMhead(vocab_size=self.vocab_size,
+                              hidden_size=self.hidden_size,
+                              dtype=dtype,
+                              rngs=self.rng,
+                              vd_sharding=(('data', 'expert', 'model'), None),
+                              dv_sharding=(None, ('data', 'expert', 'model')),
+                              random_init=force_random_weights)
+        if self.is_verbose:
+            self._print_model_architecture()
+
+    def _print_model_architecture(self):
+        num_display_layers = max(self.interleave_moe_layer_step,
+                                 self.no_rope_layer_interval)
+
+        logger.info("### Embedding ###")
+        nnx.display(self.embedder)
+
+        logger.info(f"\n### First {num_display_layers} Layers ###")
+        # Loop through the slice and display each layer
+        for i, layer in enumerate(self.layers[:num_display_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):
+        # NOTE: Since we are using nnx.eval_shape to init the model,
+        # we have to pass dynamic arrays here for __call__'s usage.
+        self.rng = nnx.Rngs(rng)
+
+        weight_loader = Llama4WeightLoader(
+            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,
+        *args,
+    ) -> Tuple[List[KVCacheType], jax.Array, List[jax.Array]]:
+        is_prefill = False
+        x_TD = self.embedder.encode(input_ids)
+
+        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
+
+
+class Llama4WeightLoader:
+
+    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.interleave_moe_layer_step = getattr(
+            vllm_config.model_config.hf_config.text_config,
+            "interleave_moe_layer_step", 1)
+
+        self.expert_prefix = "shared_expert."
+        self._transpose_map = {
+            "q_proj": (2, 0, 1),
+            "k_proj": (2, 0, 1),
+            "v_proj": (2, 0, 1),
+            "router": (1, 0),
+            f"{self.expert_prefix}down_proj": (1, 0),
+            f"{self.expert_prefix}gate_proj": (1, 0),
+            f"{self.expert_prefix}up_proj": (1, 0),
+            "feed_forward.down_proj": (1, 0),
+            "feed_forward.gate_proj": (1, 0),
+            "feed_forward.up_proj": (1, 0),
+            "o_proj": (1, 2, 0),
+            "lm_head": (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 is inverted: https://github.com/huggingface/transformers/blob/v4.53.2/src/transformers/models/llama4/modeling_llama4.py#L298
+            "o_proj": (hidden_size, attn_heads, attn_head_dim),
+        }
+
+        # Set the mappings from loaded parameter keys to standardized names.
+        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.router.weight":
+            "layers.*.custom_module.router.kernel_DE",
+            "language_model.model.layers.*.feed_forward.experts.down_proj":
+            "layers.*.custom_module.kernel_down_proj_EFD",
+            "language_model.model.layers.*.feed_forward.experts.gate_up_proj":
+            "layers.*.custom_module.kernel_up_proj_EDF",
+            "language_model.model.layers.*.feed_forward.shared_expert.down_proj.weight":
+            "layers.*.shared_experts.kernel_down_proj_FD",
+            "language_model.model.layers.*.feed_forward.shared_expert.gate_proj.weight":
+            "layers.*.shared_experts.kernel_gating_DF",
+            "language_model.model.layers.*.feed_forward.shared_expert.up_proj.weight":
+            "layers.*.shared_experts.kernel_up_proj_DF",
+            "language_model.model.layers.*.feed_forward.down_proj.weight":
+            "layers.*.custom_module.kernel_down_proj_FD",
+            "language_model.model.layers.*.feed_forward.up_proj.weight":
+            "layers.*.custom_module.kernel_up_proj_DF",
+            "language_model.model.layers.*.feed_forward.gate_proj.weight":
+            "layers.*.custom_module.kernel_gating_DF",
+        }
+
+    def map_loaded_to_standardized_name(self, loaded_key: str) -> str:
+        # Find the corresponding model key using the HF key
+        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 _map_llama4_gate_up_proj(self, model_for_loading: nnx.Module,
+                                 model_params: nnx.State, loaded_name: str,
+                                 loaded_weight: jax.Array):
+        """HF's gate_up_proj is a fused tensor of gate and up projections. It needs to be split."""
+        # gate_proj is first & up_proj is second
+        split_weights = jnp.split(loaded_weight, 2, axis=-1)
+
+        for split_type in ["gate", "up"]:
+            split_loaded_name = loaded_name.replace("gate_up_proj",
+                                                    f"{split_type}_proj")
+            if split_type == "gate":
+                mapped_name = "layers.*.custom_module.kernel_gating_EDF"
+                loaded_weight = split_weights[0]
+            else:
+                mapped_name = "layers.*.custom_module.kernel_up_proj_EDF"
+                loaded_weight = split_weights[1]
+
+            layer_num = re.search(r"layers\.(\d+)", split_loaded_name).group(1)
+            mapped_name = re.sub(r"layers\.\*", f"layers.{layer_num}",
+                                 mapped_name)
+            mapped_model_weight = get_param(model_params, mapped_name)
+
+            if mapped_model_weight.value.shape != loaded_weight.shape:
+                raise ValueError(
+                    f"Loaded shape for {split_loaded_name}: {loaded_weight.shape} "
+                    f"does not match model shape for {mapped_name}: {mapped_model_weight.value.shape}!"
+                )
+            mapped_model_weight.value = shard_put(loaded_weight,
+                                                  mapped_model_weight.sharding,
+                                                  mesh=model_for_loading.mesh)
+            logger.debug(
+                f"{split_loaded_name}: {loaded_weight.shape}  -->  {mapped_name}: {mapped_model_weight.value.shape}"
+            )
+            if self.is_verbose:
+                print_param_info(mapped_model_weight, mapped_name)
+
+    def _get_layer_num(self, loaded_key: str) -> Optional[int]:
+        """
+        Extracts the layer number from a HuggingFace weight key string.
+        Returns the layer number (int) or None if no layer number is found.
+        """
+        match = re.search(r"layers\.(\d+)", loaded_key)
+        if match:
+            return int(match.group(1))
+        return None
+
+    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:
+                is_moe_layer = False
+                layer_num = self._get_layer_num(loaded_name)
+
+                if layer_num is not None:
+                    is_moe_layer = (layer_num + 1) % \
+                            self.interleave_moe_layer_step == 0
+                    self.expert_prefix = "shared_expert." if is_moe_layer else ""
+
+                if "gate_up_proj" in loaded_name:
+                    self._map_llama4_gate_up_proj(model_for_loading,
+                                                  model_params, loaded_name,
+                                                  loaded_weight)
+                    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"):
+                    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)