feat(models): Add DBRX

server/text_generation_server/models/custom_modeling/flash_dbrx_modeling.py

@@ -35,6 +35,7 @@ from text_generation_server.utils.layers import (
     SpeculativeHead,
     get_linear,
 )
+from text_generation_server.utils.log import log_once
 
 HAS_MEGABLOCKS = True
 try:
@@ -176,33 +177,103 @@ def _load_gqa(config, prefix: str, weights):
     assert config.d_model % config.n_heads == 0
     assert config.n_heads % weights.process_group.size() == 0
 
-    weight = weights.get_weights_col_packed_qkv(
-        prefix=f"{prefix}.Wqkv",
-        quantize=config.quantize,
+    head_dim = config.d_model // config.n_heads
+    world_size = weights.process_group.size()
+    rank = weights.process_group.rank()
+
+    q_block_size = config.d_model // world_size
+    q_start = rank * q_block_size
+    q_stop = (rank + 1) * q_block_size
+
+    kv_block_size = (config.attn_config.kv_n_heads * head_dim) // world_size
+    k_offset = config.d_model
+    k_start = k_offset + rank * kv_block_size
+    k_stop = k_offset + (rank + 1) * kv_block_size
+
+    v_offset = config.d_model + config.attn_config.kv_n_heads * head_dim
+    v_start = v_offset + rank * kv_block_size
+    v_stop = v_offset + (rank + 1) * kv_block_size
+
+    if config.quantize in ["gptq", "awq"]:
+        try:
+            qweight_slice = weights._get_slice(f"{prefix}.qweight")
+            q_qweight = qweight_slice[:, q_start:q_stop]
+            k_qweight = qweight_slice[:, k_start:k_stop]
+            v_qweight = qweight_slice[:, v_start:v_stop]
+
+            qweight = torch.cat([q_qweight, k_qweight, v_qweight], dim=1)
+        except RuntimeError:
+            raise RuntimeError(
+                f"Cannot load `{config.quantize}` weight, make sure the model is already quantized"
             )
 
-    if config.quantize not in ["gptq", "awq"]:
+        qzeros_slice = weights._get_slice(f"{prefix}.qzeros")
+        q_qzeros = qzeros_slice[:, q_start:q_stop]
+        k_qzeros = qzeros_slice[:, k_start:k_stop]
+        v_qzeros = qzeros_slice[:, v_start:v_stop]
+
+        qzeros = torch.cat([q_qzeros, k_qzeros, v_qzeros], dim=1)
+
+        scales_slice = weights._get_slice(f"{prefix}.scales")
+        q_scales = scales_slice[:, q_start:q_stop]
+        k_scales = scales_slice[:, k_start:k_stop]
+        v_scales = scales_slice[:, v_start:v_stop]
+
+        scales = torch.cat([q_scales, k_scales, v_scales], dim=1)
+
+        bits, groupsize, desc_act, quant_method = weights._get_gptq_params()
+
+        from text_generation_server.utils.layers import HAS_EXLLAMA
+
+        use_exllama = (
+            bits == 4 and HAS_EXLLAMA and config.quantize == "gptq" and not desc_act
+        )
+
+        if config.quantize == "gptq" and quant_method == "gptq":
+            g_idx_slice = weights._get_slice(f"{prefix}.g_idx")
+            q_g_idx = g_idx_slice[:, q_start:q_stop]
+            k_g_idx = g_idx_slice[:, k_start:k_stop]
+            v_g_idx = g_idx_slice[:, v_start:v_stop]
+
+            w = [q_g_idx, k_g_idx, v_g_idx]
+            for w2 in w[1:]:
+                torch.testing.assert_close(w2, w[0])
+            g_idx = w[0]
+        elif config.quantize == "gptq" and quant_method == "awq":
+            log_once(
+                logger.info, "Converting AWQ model to Exllama/GPTQ packing format."
+            )
+            from text_generation_server.utils.awq.conversion_utils import (
+                fast_awq_to_gptq,
+            )
+
+            qweight, qzeros = fast_awq_to_gptq(qweight, qzeros)
+            if use_exllama:
+                g_idx = None
+            else:
+                g_idx = (
+                    torch.arange(qweight.shape[0] * (32 // bits), device=qweight.device)
+                    // groupsize
+                ).to(dtype=torch.int32)
+        else:
+            g_idx = None
+
+        weight = (qweight, qzeros, scales, g_idx, bits, groupsize, use_exllama)
+    else:
+        qkv_slice = weights._get_slice(f"{prefix}.Wqkv.weight")
+        q = qkv_slice[q_start:q_stop]
+        k = qkv_slice[k_start:k_stop]
+        v = qkv_slice[v_start:v_stop]
+
+        weight = torch.cat([q, k, v], dim=0)
         weight = weight.to(dtype=weights.dtype).to(device=weights.device)
 
-        head_size = config.d_model // config.n_heads
-        num_heads = config.n_heads // weights.process_group.size()
-        num_key_value_heads = (
-            config.attn_config.kv_n_heads // weights.process_group.size()
-        )
-        assert list(weight.shape) == [
-            (num_heads + 2 * num_key_value_heads) * head_size,
-            config.d_model,
-        ], f"{list(weight.shape)} != {[(num_heads + 2 * config.attn_config.kv_n_heads) * head_size, config.d_model]}"
-
     return TensorParallelColumnLinear(
         get_linear(weight, bias=None, quantize=config.quantize)
     )
 
 
 def _load_experts(config, prefix, weights):
-    if config.quantize is not None:
-        raise NotImplementedError("Dbrx does not support weight quantization yet.")
-
     world_size = weights.process_group.size()
     rank = weights.process_group.rank()
 
@@ -221,9 +292,9 @@ def _load_experts(config, prefix, weights):
         device=weights.device,
     )
 
-    slice_ = weights._get_slice(f"{prefix}.weight")
+    slice_ = weights._get_slice(f"{prefix}")
 
-    for i in range(config.num_local_experts):
+    for i in range(config.ffn_config.moe_num_experts):
         offset = i * expert_size
         expert_slice = slice_[start + offset : stop + offset]
 
@@ -233,6 +304,46 @@ def _load_experts(config, prefix, weights):
     return tensor
 
 
+def _load_experts_quantized(config, prefix, weights, cls):
+    world_size = weights.process_group.size()
+    rank = weights.process_group.rank()
+
+    assert (
+        config.ffn_config.ffn_hidden_size % world_size == 0
+    ), f"The chosen size {config.ffn_config.ffn_hidden_size} is not compatible with sharding on {world_size} shards"
+
+    expert_size = config.ffn_config.ffn_hidden_size
+    block_size = expert_size // world_size
+    start = rank * block_size
+    stop = (rank + 1) * block_size
+
+    slice_ = weights._get_slice(f"{prefix}")
+
+    experts = []
+    for i in range(config.ffn_config.moe_num_experts):
+        if config.quantize in ["gptq", "awq"]:
+            raise NotImplementedError(
+                "Dbrx does not support gptq/awq quantization yet."
+            )
+        else:
+            offset = i * expert_size
+            expert_slice = (
+                slice_[start + offset : stop + offset]
+                .to(dtype=weights.dtype)
+                .to(device=weights.device)
+            )
+
+        if cls == TensorParallelRowLinear:
+            expert_slice = expert_slice.t().contiguous()
+            linear = get_linear(expert_slice, None, config.quantize)
+            experts.append(cls(linear, weights.process_group))
+        else:
+            linear = get_linear(expert_slice, None, config.quantize)
+            experts.append(cls(linear))
+
+    return experts
+
+
 class DbrxAttention(torch.nn.Module):
     def __init__(
         self,
@@ -391,9 +502,7 @@ class DbrxNormAttentionNorm(nn.Module):
         )
 
         # faster post attention rms norm
-        normed_attn_res_output, attn_res = self.post_attention_layernorm(
-            attn_output, res
-        )
+        normed_attn_res_output, attn_res = self.norm_2(attn_output, res)
 
         return normed_attn_res_output, attn_res
 
@@ -663,6 +772,7 @@ class BlockSparseMoE(nn.Module):
             weights = weights / torch.norm(
                 weights, p=self.moe_normalize_expert_weights, dim=-1, keepdim=True
             )
+        weights = weights.to(x.dtype)
 
         # Expand to [num_experts, sequence_length, model_dim]
         x = x.view(1, -1, input_shape[-1]).expand(self.num_experts, -1, input_shape[-1])
@@ -703,8 +813,6 @@ class DenseMoE(nn.Module):
     def __init__(self, prefix, config: DbrxConfig, weights):
         super().__init__()
 
-        raise NotImplementedError("Quantization is not implemented for Dbrx")
-
         self.moe_normalize_expert_weights = (
             config.ffn_config.moe_normalize_expert_weights
         )
@@ -731,24 +839,24 @@ class DenseMoE(nn.Module):
             config, f"{prefix}.router.layer", weights, bias=False
         )
 
-        self.w1 = [
-            TensorParallelColumnLinear.load(
-                config, prefix=f"{prefix}.experts.{i}.w1", weights=weights, bias=False
+        self.w1 = _load_experts_quantized(
+            config,
+            prefix=f"{prefix}.experts.mlp.w1",
+            weights=weights,
+            cls=TensorParallelColumnLinear,
         )
-            for i in range(self.num_experts)
-        ]
-        self.w3 = [
-            TensorParallelColumnLinear.load(
-                config, prefix=f"{prefix}.experts.{i}.w3", weights=weights, bias=False
+        self.w2 = _load_experts_quantized(
+            config,
+            prefix=f"{prefix}.experts.mlp.w2",
+            weights=weights,
+            cls=TensorParallelRowLinear,
         )
-            for i in range(self.num_experts)
-        ]
-        self.w2 = [
-            TensorParallelRowLinear.load(
-                config, prefix=f"{prefix}.experts.{i}.w2", weights=weights, bias=False
+        self.v1 = _load_experts_quantized(
+            config,
+            prefix=f"{prefix}.experts.mlp.v1",
+            weights=weights,
+            cls=TensorParallelColumnLinear,
         )
-            for i in range(self.num_experts)
-        ]
 
         self.process_group = weights.process_group
 
@@ -764,26 +872,30 @@ class DenseMoE(nn.Module):
         # gate_logits: (sequence_length, n_experts)
         gate_logits = self.gate(x)
         # all_probs: (sequence_length, n_experts) and upcast for softmax
-        all_probs = torch.nn.functional.softmax(gate_logits, dim=1, dtype=torch.float)
+        weights = torch.nn.functional.softmax(gate_logits, dim=1, dtype=torch.float)
 
         if self.top_k < self.num_experts:
             _, not_selected_experts = torch.topk(
-                all_probs,
+                weights,
                 self.num_experts - self.top_k,
                 largest=False,
                 sorted=False,
                 dim=1,
             )
             # Mask not selected experts
-            all_probs.scatter_(1, not_selected_experts, 0)
+            weights.scatter_(1, not_selected_experts, 0)
 
         # Re-normalize
-        weights = all_probs / all_probs.sum(dim=1, keepdim=True)
+        if self.moe_normalize_expert_weights:
+            weights = weights / torch.norm(
+                weights, p=self.moe_normalize_expert_weights, dim=-1, keepdim=True
+            )
+        weights = weights.to(x.dtype)
 
         # Final output tensor
         out = x.new_zeros(x.shape[0], self.hidden_dim)
         for i in range(self.num_experts):
-            h = self.act(self.w1[i](x)) * self.w3[i](x)
+            h = self.act(self.w1[i](x)) * self.v1[i](x)
             h = self.w2[i](h, reduce=False)
             # Add expert output to out with masking
             out += h * weights[:, i].view(-1, 1)
@@ -821,7 +933,7 @@ class DbrxLayer(nn.Module):
         max_s,
     ):
         # Self Attention
-        attn_output, attn_res = self.self_attn(
+        attn_output, attn_res = self.attn(
             hidden_states,
             residual,
             cos,
@@ -861,9 +973,9 @@ class DbrxModel(torch.nn.Module):
             prefix="transformer.norm_f", weights=weights, eps=1e-5
         )
 
-        self.head_size = self.layers[0].self_attn.head_size
-        self.num_heads = self.layers[0].self_attn.num_heads
-        self.num_key_value_heads = self.layers[0].self_attn.num_key_value_heads
+        self.head_size = self.layers[0].attn.self_attn.head_size
+        self.num_heads = self.layers[0].attn.self_attn.num_heads
+        self.num_key_value_heads = self.layers[0].attn.self_attn.num_key_value_heads
 
     def forward(
         self,
@@ -880,7 +992,7 @@ class DbrxModel(torch.nn.Module):
 
         # Get rotary cos and sin for this forward
         # Avoid to index in each layer
-        cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
+        cos, sin = self.layers[0].attn.self_attn.rotary_emb.get_cos_sin(
             position_ids, max_s, hidden_states.dtype
         )
 

server/text_generation_server/models/custom_modeling/flash_mixtral_modeling.py

@@ -552,6 +552,7 @@ class BlockSparseMoE(nn.Module):
 
         # Re-normalize
         weights = all_probs / all_probs.sum(dim=1, keepdim=True)
+        weights = weights.to(x.dtype)
 
         # Expand to [num_experts, sequence_length, model_dim]
         x = x.view(1, -1, input_shape[-1]).expand(self.num_experts, -1, input_shape[-1])
@@ -660,6 +661,7 @@ class DenseMoE(nn.Module):
 
         # Re-normalize
         weights = all_probs / all_probs.sum(dim=1, keepdim=True)
+        weights = weights.to(x.dtype)
 
         # Final output tensor
         out = x.new_zeros(x.shape[0], self.hidden_dim)

server/text_generation_server/models/flash_dbrx.py

@@ -3,6 +3,7 @@ import torch.distributed
 
 from opentelemetry import trace
 from typing import Optional
+from transformers import AutoTokenizer
 from transformers.models.gpt2 import GPT2TokenizerFast
 
 from text_generation_server.models import FlashCausalLM
@@ -36,6 +37,17 @@ class FlashDbrx(FlashCausalLM):
         else:
             raise NotImplementedError("FlashDBRX is only available on GPU")
 
+        try:
+            tokenizer = AutoTokenizer.from_pretrained(
+                model_id,
+                revision=revision,
+                padding_side="left",
+                truncation_side="left",
+                trust_remote_code=trust_remote_code,
+                use_fast=True,
+                from_slow=False,
+            )
+        except:
             # FIXME: change back to model id once the tokenizer.json is merged
             tokenizer = GPT2TokenizerFast.from_pretrained(
                 "Xenova/dbrx-instruct-tokenizer",