chore: import upstream snapshot with attribution

python/mlc_llm/model/internlm2/internlm2_loader.py

@@ -0,0 +1,94 @@
+"""
+This file specifies how MLC's InternLM2 parameter maps from other formats, for example HuggingFace
+PyTorch, HuggingFace safetensors.
+"""
+
+import functools
+
+import numpy as np
+
+from mlc_llm.loader import ExternMapping
+from mlc_llm.quantization import Quantization
+
+from .internlm2_model import InternLM2ForCausalLM
+
+
+def huggingface(model_config: InternLM2ForCausalLM, quantization: Quantization) -> ExternMapping:
+    """Returns a parameter mapping that maps from the names of MLC LLM parameters to
+    the names of HuggingFace PyTorch parameters.
+
+    Parameters
+    ----------
+    model_config : InternLM2Config
+        The configuration of the InternLM2 model.
+
+    quantization : Quantization
+        The quantization configuration.
+
+    Returns
+    -------
+    param_map : ExternMapping
+        The parameter mapping from MLC to HuggingFace PyTorch.
+    """
+    model = InternLM2ForCausalLM(model_config)
+    if quantization is not None:
+        model.to(quantization.model_dtype)
+    _, _named_params, _ = model.export_tvm(
+        spec=model.get_default_spec(),
+        allow_extern=True,
+    )
+    named_parameters = dict(_named_params)
+
+    mapping = ExternMapping()
+
+    def _convert_wqkv_layout(wqkv, dtype):
+        config = model_config
+        kv_groups = config.num_attention_heads // config.num_key_value_heads
+        head_dim = config.hidden_size // config.num_attention_heads
+        wqkv = wqkv.reshape(-1, 2 + kv_groups, head_dim, wqkv.shape[-1])
+        wq, wk, wv = np.split(wqkv, [kv_groups, kv_groups + 1], axis=1)
+        wq = wq.reshape(-1, wq.shape[-1])
+        wk = wk.reshape(-1, wk.shape[-1])
+        wv = wv.reshape(-1, wv.shape[-1])
+        return np.concatenate([wq, wk, wv], axis=0).astype(dtype)
+
+    for i in range(model_config.num_hidden_layers):
+        # Add gates in MLP
+        mlp = f"model.layers.{i}.feed_forward"
+        mlc_name = f"{mlp}.gate_up_proj.weight"
+        mlc_param = named_parameters[mlc_name]
+        mapping.add_mapping(
+            mlc_name,
+            [
+                f"{mlp}.w1.weight",
+                f"{mlp}.w3.weight",
+            ],
+            functools.partial(
+                lambda w1, w3, dtype: np.concatenate([w1, w3], axis=0).astype(dtype),
+                dtype=mlc_param.dtype,
+            ),
+        )
+
+        mlc_name = f"model.layers.{i}.attention.wqkv.weight"
+        mlc_param = named_parameters[mlc_name]
+        mapping.add_mapping(
+            mlc_name,
+            [mlc_name],
+            functools.partial(
+                _convert_wqkv_layout,
+                dtype=mlc_param.dtype,
+            ),
+        )
+
+    for mlc_name, mlc_param in named_parameters.items():
+        if mlc_name not in mapping.param_map:
+            mapping.add_mapping(
+                mlc_name,
+                [mlc_name],
+                functools.partial(
+                    lambda x, dtype: x.astype(dtype),
+                    dtype=mlc_param.dtype,
+                ),
+            )
+
+    return mapping

python/mlc_llm/model/internlm2/internlm2_model.py

@@ -0,0 +1,377 @@
+"""
+Implementation for InternLM2 architecture.
+"""
+
+import dataclasses
+from typing import Any, Dict, Optional  # noqa: UP035
+
+from tvm import tirx
+from tvm.relax.frontend import nn
+from tvm.relax.frontend.nn import Tensor, op
+
+from mlc_llm import op as op_ext
+from mlc_llm.model.model_utils import index_last_token
+from mlc_llm.nn import PagedKVCache, RopeMode
+from mlc_llm.support import logging
+from mlc_llm.support import tensor_parallel as tp
+from mlc_llm.support.config import ConfigBase
+from mlc_llm.support.style import bold
+
+logger = logging.getLogger(__name__)
+
+
+@dataclasses.dataclass
+class InternLM2Config(ConfigBase):
+    """Configuration of the InternLM2 model."""
+
+    vocab_size: int
+    hidden_size: int
+    num_hidden_layers: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    rms_norm_eps: float
+    intermediate_size: int
+    bias: bool
+    use_cache: bool
+    rope_theta: int
+    pad_token_id: int
+    bos_token_id: int
+    eos_token_id: int
+    context_window_size: int = 0
+    prefill_chunk_size: int = 0
+    tensor_parallel_shards: int = 1
+    max_batch_size: int = 1
+    head_dim: int = 0
+    kwargs: Dict[str, Any] = dataclasses.field(default_factory=dict)  # noqa: UP006
+
+    def __post_init__(self):
+        if self.context_window_size == 0:
+            for name in ["max_position_embeddings", "max_sequence_length"]:
+                if name in self.kwargs:
+                    self.context_window_size = self.kwargs.pop(name)
+                    logger.info(
+                        "%s not found in config.json. Falling back to %s (%d)",
+                        bold("context_window_size"),
+                        bold(name),
+                        self.context_window_size,
+                    )
+                    break
+            else:
+                raise ValueError(
+                    "Unable to determine the maximum sequence length, because none of "
+                    "`context_window_size`, `max_position_embeddings` or `max_sequence_length` is "
+                    "provided in `config.json`."
+                )
+        if self.head_dim == 0:
+            self.head_dim = self.hidden_size // self.num_attention_heads
+        assert self.head_dim * self.num_attention_heads == self.hidden_size
+        if self.prefill_chunk_size == 0:
+            logger.info(
+                "%s defaults to %d",
+                bold("prefill_chunk_size"),
+                min(self.context_window_size, 2048),
+            )
+            self.prefill_chunk_size = min(self.context_window_size, 2048)
+        elif self.prefill_chunk_size > self.context_window_size:
+            logger.info(
+                "Overriding %s from %d to %d",
+                bold("prefill_chunk_size"),
+                self.prefill_chunk_size,
+                min(self.context_window_size, 2048),
+            )
+            self.prefill_chunk_size = min(self.context_window_size, 2048)
+
+
+class InternLM2Attention(nn.Module):
+    def __init__(self, config: InternLM2Config):
+        if config.num_attention_heads % config.tensor_parallel_shards != 0:
+            raise ValueError(
+                f"Cannot split {config.num_attention_heads} attention heads "
+                f"evenly to {config.tensor_parallel_shards} GPUs."
+            )
+        self.hidden_size = config.hidden_size
+        self.rope_theta = config.rope_theta
+        self.num_heads = config.num_attention_heads // config.tensor_parallel_shards
+        self.head_dim = config.head_dim
+        self.num_key_value_heads = config.num_key_value_heads // config.tensor_parallel_shards
+        self.max_position_embeddings = config.context_window_size
+
+        self.wqkv = nn.Linear(
+            self.hidden_size,
+            (self.num_heads + 2 * self.num_key_value_heads) * self.head_dim,
+            bias=config.bias,
+        )
+        self.wo = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.bias)
+
+    def forward(self, hidden_states: Tensor, paged_kv_cache: PagedKVCache, layer_id: int):
+        d, h_q, h_kv = self.head_dim, self.num_heads, self.num_key_value_heads
+        b, s, _ = hidden_states.shape
+        qkv = self.wqkv(hidden_states)
+        qkv = op.reshape(qkv, (b, s, h_q + h_kv + h_kv, d))
+        output = op.reshape(
+            paged_kv_cache.attention_with_fused_qkv(
+                layer_id, qkv, self.num_heads, sm_scale=self.head_dim**-0.5
+            ),
+            (b, s, h_q * d),
+        )
+        attn_output = self.wo(output)
+        return attn_output
+
+
+class InternLM2MLP(nn.Module):
+    def __init__(self, config: InternLM2Config):
+        if config.intermediate_size % config.tensor_parallel_shards != 0:
+            raise ValueError(
+                f"Cannot split MLP intermediate size {config.intermediate_size} "
+                f"evenly to {config.tensor_parallel_shards} GPUs."
+            )
+        self.intermediate_size = config.intermediate_size // config.tensor_parallel_shards
+        self.gate_up_proj = nn.Linear(
+            in_features=config.hidden_size,
+            out_features=2 * self.intermediate_size,
+            bias=False,
+        )
+        self.w2 = nn.Linear(self.intermediate_size, config.hidden_size, bias=False)
+
+    def forward(self, x: Tensor):
+        concat_x1_x2 = self.gate_up_proj(x)
+        x1, x2 = op.split(concat_x1_x2, 2, axis=-1)
+        return self.w2(op.silu(x1) * x2)
+
+
+class InternLM2DecoderLayer(nn.Module):
+    def __init__(self, config: InternLM2Config):
+        self.attention = InternLM2Attention(config)
+        self.feed_forward = InternLM2MLP(config)
+        self.attention_norm = nn.RMSNorm(config.hidden_size, -1, config.rms_norm_eps, bias=False)
+        self.ffn_norm = nn.RMSNorm(config.hidden_size, -1, config.rms_norm_eps, bias=False)
+
+        def _set_tp():
+            def _set(layer, hint):
+                layer.attrs["shard_strategy"] = hint
+
+            hd = config.head_dim
+            q = self.attention.num_heads * hd
+            k = self.attention.num_key_value_heads * hd
+            v = self.attention.num_key_value_heads * hd
+            i = self.feed_forward.intermediate_size
+            _set(
+                self.attention.wqkv.weight,
+                tp.ShardSingleDim("_shard_qkv_weight", dim=0, segs=[q, k, v]),
+            )
+            if config.bias:
+                _set(
+                    self.attention.wqkv.bias,
+                    tp.ShardSingleDim("_shard_qkv_bias", dim=0, segs=[q, k, v]),
+                )
+            _set(self.attention.wo.weight, tp.ShardSingleDim("_shard_o", dim=1))
+            _set(
+                self.feed_forward.gate_up_proj.weight,
+                tp.ShardSingleDim("_shard_mlp_up", segs=[i, i], dim=0),
+            )
+            _set(self.feed_forward.w2.weight, tp.ShardSingleDim("_shard_mlp_down", dim=1))
+
+        self.tensor_parallel_shards = config.tensor_parallel_shards
+        _set_tp()
+
+    def forward(self, hidden_states: Tensor, paged_kv_cache: PagedKVCache, layer_id: int):
+        residual = hidden_states
+        hidden_states = self.attention_norm(hidden_states)
+        hidden_states = self.attention(hidden_states, paged_kv_cache, layer_id)
+        hidden_states = self._apply_residual(hidden_states, residual=residual)
+        residual = hidden_states
+        hidden_states = self.ffn_norm(hidden_states)
+        hidden_states = self.feed_forward(hidden_states)
+        hidden_states = self._apply_residual(hidden_states, residual=residual)
+        return hidden_states
+
+    def _apply_residual(self, out, residual):
+        if self.tensor_parallel_shards > 1:
+            return op.ccl_allreduce(out, "sum") + residual
+        return out + residual
+
+
+class InternLM2Model(nn.Module):
+    def __init__(self, config: InternLM2Config):
+        self.padding_idx = config.pad_token_id
+        self.tok_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList(
+            [InternLM2DecoderLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.norm = nn.RMSNorm(config.hidden_size, -1, config.rms_norm_eps, bias=False)
+
+    def forward(self, inputs: Tensor, paged_kv_cache: PagedKVCache):
+        hidden_states = inputs
+        for layer_id, layer in enumerate(self.layers):
+            hidden_states = layer(hidden_states, paged_kv_cache, layer_id)
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+class InternLM2ForCausalLM(nn.Module):
+    def __init__(self, config: InternLM2Config):
+        self.model = InternLM2Model(config)
+        self.output = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.vocab_size = config.vocab_size
+        self.dtype = "float32"
+        self.num_hidden_layers = config.num_hidden_layers
+        self.hidden_size = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.head_dim = config.head_dim
+        self.rope_theta = config.rope_theta
+        self.tensor_parallel_shards = config.tensor_parallel_shards
+
+    def to(self, dtype: Optional[str] = None):
+        super().to(dtype=dtype)
+        if dtype is not None:
+            self.dtype = dtype
+
+    def batch_forward(
+        self,
+        input_embeds: Tensor,
+        paged_kv_cache: PagedKVCache,
+        logit_positions: Optional[Tensor] = None,
+    ):
+        op_ext.configure()
+
+        hidden_states = self.model(input_embeds, paged_kv_cache)
+        if logit_positions is not None:
+            hidden_states = op.take(hidden_states, logit_positions, axis=1)
+        logits = self.output(hidden_states)
+        if logits.dtype != "float32":
+            logits = logits.astype("float32")
+        return logits
+
+    def embed(self, input_ids: Tensor):
+        if self.tensor_parallel_shards > 1:
+            input_ids = op.ccl_broadcast_from_worker0(input_ids)
+        return self.model.tok_embeddings(input_ids)
+
+    def prefill(self, input_embed: Tensor, paged_kv_cache: PagedKVCache):
+        op_ext.configure()
+
+        hidden_states = self.model(input_embed, paged_kv_cache)
+        hidden_states = index_last_token(hidden_states)
+        logits = self.output(hidden_states)
+        if logits.dtype != "float32":
+            logits = logits.astype("float32")
+        return logits, paged_kv_cache
+
+    def decode(self, input_embed: Tensor, paged_kv_cache: PagedKVCache):
+        op_ext.configure()
+
+        hidden_states = self.model(input_embed, paged_kv_cache)
+        logits = self.output(hidden_states)
+        if logits.dtype != "float32":
+            logits = logits.astype("float32")
+        return logits, paged_kv_cache
+
+    def batch_prefill(
+        self,
+        input_embeds: Tensor,
+        logit_positions: Tensor,
+        paged_kv_cache: PagedKVCache,
+    ):
+        if self.tensor_parallel_shards > 1:
+            logit_positions = op.ccl_broadcast_from_worker0(logit_positions)
+        logits = self.batch_forward(input_embeds, paged_kv_cache, logit_positions)
+        return logits, paged_kv_cache
+
+    def batch_decode(self, input_embeds: Tensor, paged_kv_cache: PagedKVCache):
+        logits = self.batch_forward(input_embeds, paged_kv_cache)
+        return logits, paged_kv_cache
+
+    def batch_verify(self, input_embeds: Tensor, paged_kv_cache: PagedKVCache):
+        logits = self.batch_forward(input_embeds, paged_kv_cache)
+        return logits, paged_kv_cache
+
+    def create_paged_kv_cache(
+        self,
+        max_batch_size: tirx.Var,
+        max_total_seq_len: tirx.Var,
+        prefill_chunk_size: tirx.Var,
+        page_size: tirx.Var,
+        support_sliding_window: tirx.Var,
+    ) -> PagedKVCache:
+        return PagedKVCache.create_generic(
+            attn_kind="mha",
+            max_batch_size=max_batch_size,
+            max_total_seq_len=max_total_seq_len,
+            prefill_chunk_size=prefill_chunk_size,
+            page_size=page_size,
+            support_sliding_window=support_sliding_window,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads // self.tensor_parallel_shards,
+            num_key_value_heads=self.num_key_value_heads // self.tensor_parallel_shards,
+            qk_head_dim=self.head_dim,
+            v_head_dim=self.head_dim,
+            rope_mode=RopeMode.NORMAL,
+            rope_scale=1,
+            rope_theta=self.rope_theta,
+            dtype=self.dtype,
+        )
+
+    def get_default_spec(self):
+        mod_spec = {
+            "embed": {
+                "input_ids": nn.spec.Tensor(["seq_len"], "int32"),
+                "$": {
+                    "param_mode": "packed",
+                    "effect_mode": "none",
+                },
+            },
+            "prefill": {
+                "input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
+                "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
+                "$": {
+                    "param_mode": "packed",
+                    "effect_mode": "none",
+                },
+            },
+            "decode": {
+                "input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
+                "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
+                "$": {
+                    "param_mode": "packed",
+                    "effect_mode": "none",
+                },
+            },
+            "batch_prefill": {
+                "input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
+                "logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
+                "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
+                "$": {
+                    "param_mode": "packed",
+                    "effect_mode": "none",
+                },
+            },
+            "batch_decode": {
+                "input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
+                "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
+                "$": {
+                    "param_mode": "packed",
+                    "effect_mode": "none",
+                },
+            },
+            "batch_verify": {
+                "input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
+                "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
+                "$": {
+                    "param_mode": "packed",
+                    "effect_mode": "none",
+                },
+            },
+            "create_paged_kv_cache": {
+                "max_batch_size": int,
+                "max_total_seq_len": int,
+                "prefill_chunk_size": int,
+                "page_size": int,
+                "support_sliding_window": int,
+                "$": {
+                    "param_mode": "none",
+                    "effect_mode": "none",
+                },
+            },
+        }
+        return nn.spec.ModuleSpec.from_raw(mod_spec, self)