[SLM] GPTJ Multi-GPU support (#3070)

This PR supports TP function of GPTJ Model and fix minor typo of OlMo Model.

Author: tlopex

python/mlc_llm/model/gpt_j/gpt_j_model.py

@@ -13,6 +13,7 @@
 from mlc_llm import op as op_ext
 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
 
@@ -57,6 +58,9 @@ def __post_init__(self):
                     "`context_window_size`, `max_position_embeddings` or `max_sequence_length` is "
                     "provided in `config.json`."
                 )
+        if self.head_dim == 0:
+            self.head_dim = self.n_embd // self.n_head
+        assert self.head_dim * self.n_head == self.n_embd
         if self.prefill_chunk_size == 0:
             logger.info(
                 "%s defaults to %d",
@@ -72,7 +76,6 @@ def __post_init__(self):
                 min(self.context_window_size, 8192),
             )
             self.prefill_chunk_size = min(self.context_window_size, 8192)
-        assert self.tensor_parallel_shards == 1, "GPTJ currently does not support sharding."
 
 
 # pylint: disable=invalid-name,missing-docstring
@@ -82,7 +85,7 @@ class GPTJAttention(nn.Module):  # pylint: disable=too-many-instance-attributes
     def __init__(self, config: GPTJConfig):
         self.embed_dim = config.n_embd
         self.num_heads = config.n_head // config.tensor_parallel_shards
-        self.head_dim = self.embed_dim // self.num_heads
+        self.head_dim = config.head_dim
         self.max_position_embeddings = config.context_window_size
         self.rope_theta = 10000
         self.rotary_dim = config.rotary_dim
@@ -140,14 +143,41 @@ def __init__(self, config: GPTJConfig):
         self.attn = GPTJAttention(config)
         self.mlp = GPTJMLP(config)
 
+        def _set_tp():
+            def _set(layer, hint):
+                layer.attrs["shard_strategy"] = hint
+
+            hd = config.head_dim
+            q = self.attn.num_heads * hd
+            k = self.attn.num_heads * hd
+            v = self.attn.num_heads * hd
+            _set(
+                self.attn.c_attn.weight,
+                tp.ShardSingleDim("_shard_qkv_weight", dim=0, segs=[q, k, v]),
+            )
+            _set(self.attn.out_proj.weight, tp.ShardSingleDim("_shard_o", dim=1))
+            _set(
+                self.mlp.fc_in.weight,
+                tp.ShardSingleDim("_shard_c_fc_weight", dim=0),
+            )
+            _set(self.mlp.fc_out.weight, tp.ShardSingleDim("_shard_mlp_c_proj", 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.ln_1(hidden_states)
         attn_output = self.attn(hidden_states, paged_kv_cache, layer_id)
         feed_forward_hidden_states = self.mlp(hidden_states)
-        hidden_states = attn_output + feed_forward_hidden_states + residual
+        hidden_states = self._apply_residual(attn_output + feed_forward_hidden_states, 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 GPTJModel(nn.Module):
     def __init__(self, config: GPTJConfig):

python/mlc_llm/model/olmo/olmo_model.py

@@ -102,7 +102,7 @@ def __post_init__(self):  # pylint: disable=too-many-branches
                 raise ValueError(f"'clip_qkv'({self.clip_qkv}) should be non-negative")
 
 
-class OLMoEebedding(nn.Embedding):
+class OLMoEmbedding(nn.Embedding):
     """The embedding module that can be shared with the final lm_head. From Qwen2Embedding."""
 
     def lm_head_forward(self, x: nn.Tensor):
@@ -248,7 +248,7 @@ def forward(  # pylint: disable=missing-function-docstring
 class OLMoModel(nn.Module):  # pylint: disable=missing-class-docstring
     def __init__(self, config: OLMoConfig):
         assert config.hidden_size % config.num_attention_heads == 0
-        self.embed_tokens = OLMoEebedding(config.vocab_size, config.hidden_size)
+        self.embed_tokens = OLMoEmbedding(config.vocab_size, config.hidden_size)
         self.layers = nn.ModuleList(
             [OLMoDecoderLayer(config) for _ in range(config.num_hidden_layers)]
         )