support qwen3 on nvidia

server/text_generation_server/models/__init__.py

@@ -152,6 +152,9 @@ try:
     from text_generation_server.models.custom_modeling.flash_qwen2_modeling import (
         Qwen2ForCausalLM,
     )
+    from text_generation_server.models.custom_modeling.flash_qwen3_modeling import (
+        Qwen3ForCausalLM,
+    )
     from text_generation_server.models.custom_modeling.flash_mistral_modeling import (
         FlashMistralForCausalLM,
     )
@@ -348,6 +351,11 @@ class ModelType(enum.Enum):
         "name": "Qwen 2",
         "url": "https://huggingface.co/collections/Qwen/qwen2-6659360b33528ced941e557f",
     }
+    QWEN3 = {
+        "type": "qwen3",
+        "name": "Qwen 3",
+        "url": "https://huggingface.co/collections/Qwen/qwen3-67c6c6f89c4f76621268bb6d",
+    }
     QWEN2_VL = {
         "type": "qwen2_vl",
         "name": "Qwen 2 VL",
@@ -1470,6 +1478,40 @@ def get_model(
                 trust_remote_code=trust_remote_code,
             )
 
+    if model_type == QWEN3:
+        if FLASH_ATTENTION:
+            return FlashCausalLM(
+                model_id=model_id,
+                model_class=Qwen3ForCausalLM,
+                revision=revision,
+                quantize=quantize,
+                speculator=speculator,
+                dtype=dtype,
+                kv_cache_dtype=kv_cache_dtype,
+                trust_remote_code=trust_remote_code,
+                lora_adapter_ids=lora_adapter_ids,
+            )
+        elif FLASH_TRANSFORMERS_BACKEND:
+            return TransformersFlashCausalLM.fallback(
+                model_id,
+                revision,
+                quantize=quantize,
+                speculator=speculator,
+                dtype=dtype,
+                trust_remote_code=trust_remote_code,
+            )
+        elif sharded:
+            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Qwen3"))
+        else:
+            return CausalLM.fallback(
+                model_id,
+                revision,
+                quantize=quantize,
+                speculator=speculator,
+                dtype=dtype,
+                trust_remote_code=trust_remote_code,
+            )
+
     if model_type == OPT:
         return CausalLM(
             model_id=model_id,

server/text_generation_server/models/custom_modeling/flash_qwen3_modeling.py

@@ -0,0 +1,464 @@
+import torch
+import torch.distributed
+
+from torch import nn
+from transformers.activations import ACT2FN
+from typing import Optional, List, Tuple
+
+from text_generation_server.layers.attention import (
+    paged_attention,
+    attention,
+    Seqlen,
+)
+from text_generation_server.layers import (
+    TensorParallelMultiAdapterLinear,
+    TensorParallelAdapterRowLinear,
+    TensorParallelRowLinear,
+    TensorParallelColumnLinear,
+    TensorParallelEmbedding,
+    SpeculativeHead,
+)
+from text_generation_server.layers.attention.kv_cache import get_kv_scales
+from text_generation_server.layers.rotary import PositionRotaryEmbedding
+from text_generation_server.layers.layernorm import (
+    FastRMSNorm,
+)
+
+
+def load_attention(config, prefix, weights, layer_id):
+    prefixes = [f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"]
+    head_size = getattr(
+        config, "head_dim", config.hidden_size // config.num_attention_heads
+    )
+    sizes = [
+        head_size * config.num_attention_heads,
+        head_size * config.num_key_value_heads,
+        head_size * config.num_key_value_heads,
+    ]
+    if config.num_attention_heads != config.num_key_value_heads:
+        base_layer = _load_gqa(config, prefix, weights)
+    else:
+        base_layer = TensorParallelColumnLinear.load_multi(
+            config,
+            prefixes=prefixes,
+            dim=0,
+            weights=weights,
+            bias=getattr(config, 'attention_bias', False),  # Use config value like vLLM
+        )
+    return TensorParallelMultiAdapterLinear.load(
+        base_layer=base_layer,
+        layer_id=layer_id,
+        layer_names=prefixes,
+        sizes=sizes,
+        process_group=weights.process_group,
+    )
+
+
+def _load_gqa(config, prefix: str, weights):
+    assert config.hidden_size % config.num_attention_heads == 0
+    assert config.num_attention_heads % weights.process_group.size() == 0
+
+    return TensorParallelColumnLinear.load_multi(
+        config,
+        prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"],
+        dim=0,
+        weights=weights,
+        bias=getattr(config, 'attention_bias', False),  # Use config value like vLLM
+    )
+
+
+class Qwen3Attention(torch.nn.Module):
+    def __init__(
+        self,
+        index: int,
+        prefix: str,
+        config,
+        weights,
+    ):
+        super().__init__()
+        self.layer_idx = index
+        self.config = config
+        self.head_dim = getattr(
+            config, "head_dim", config.hidden_size // config.num_attention_heads
+        )
+        self.num_key_value_groups = (
+            config.num_attention_heads // config.num_key_value_heads
+        )
+        self.num_heads = config.num_attention_heads
+        self.attention_dropout = config.attention_dropout
+        self.softmax_scale = self.head_dim**-0.5
+
+        self.window_size = (
+            config.sliding_window if config.sliding_window is not None else -1
+        )
+        # Handle sliding window configuration similar to Intel Gaudi version
+        self.sliding_window = config.sliding_window
+        if hasattr(config, 'use_sliding_window') and hasattr(config, 'max_window_layers'):
+            if not (
+                config.use_sliding_window
+                and getattr(config, "sliding_window", None) is not None
+                and self.layer_idx >= config.max_window_layers
+            ):
+                self.sliding_window = None
+
+        self.rotary_emb = PositionRotaryEmbedding.static(
+            config=config,
+            dim=self.head_dim,
+            base=config.rope_theta,
+            device=weights.device,
+        )
+
+        if self.num_heads % weights.process_group.size() != 0:
+            raise ValueError(
+                f"`num_heads` must be divisible by `num_shards` (got `num_heads`: {self.num_heads} "
+                f"and `num_shards`: {weights.process_group.size()}"
+            )
+        self.num_heads = self.num_heads // weights.process_group.size()
+        self.num_key_value_heads = (
+            config.num_key_value_heads // weights.process_group.size()
+        )
+
+        self.query_key_value = load_attention(config, prefix, weights, index)
+
+        self.kv_scales = get_kv_scales(weights, f"{prefix}")
+
+        # Q and K normalization layers
+        self.q_norm = FastRMSNorm.load(
+            prefix=f"{prefix}.q_norm",
+            weights=weights,
+            eps=config.rms_norm_eps,
+        )
+        self.k_norm = FastRMSNorm.load(
+            prefix=f"{prefix}.k_norm",
+            weights=weights,
+            eps=config.rms_norm_eps,
+        )
+
+        o_proj = TensorParallelRowLinear.load(
+            config,
+            prefix=f"{prefix}.o_proj",
+            weights=weights,
+            bias=False,
+        )
+        self.o_proj = TensorParallelAdapterRowLinear.load(
+            o_proj,
+            index,
+            "o_proj",
+            process_group=weights.process_group,
+        )
+        self.num_groups = self.num_heads // self.num_key_value_heads
+        self.kv_head_mapping = torch.arange(
+            0, self.num_key_value_heads, dtype=torch.int32, device=weights.device
+        ).repeat_interleave(self.num_groups)
+
+    def forward(
+        self,
+        hidden_states,
+        cos,
+        sin,
+        cu_seqlen_prefill,
+        kv_cache,
+        block_tables,
+        slots,
+        seqlen,
+        max_s,
+        prefill_cache_indices,
+        adapter_data,
+    ):
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        qkv = self.query_key_value(hidden_states, adapter_data)
+        query_states, key_states, value_states = qkv.split(
+            [
+                self.head_dim * self.num_heads,
+                self.head_dim * self.num_key_value_heads,
+                self.head_dim * self.num_key_value_heads,
+            ],
+            dim=1,
+        )
+
+        # First reshape to head dimensions
+        query_states = query_states.reshape(hidden_shape)
+        key_states = key_states.reshape(hidden_shape)
+        value_states = value_states.reshape(hidden_shape)
+
+        # Apply Q and K normalization on head_dim - following vLLM/SGLang correct pattern
+        # This matches the reference implementations and is the correct approach
+        q_by_head = query_states.reshape(-1, self.head_dim)
+        q_by_head, _ = self.q_norm(q_by_head)
+        query_states = q_by_head.view(query_states.shape)
+        
+        k_by_head = key_states.reshape(-1, self.head_dim)
+        k_by_head, _ = self.k_norm(k_by_head)
+        key_states = k_by_head.view(key_states.shape)
+
+        self.rotary_emb(query_states, key_states, cos, sin)
+
+        if prefill_cache_indices is not None:
+            key_to_cache = key_states[prefill_cache_indices]
+            value_to_cache = value_states[prefill_cache_indices]
+        else:
+            key_to_cache = key_states
+            value_to_cache = value_states
+
+        kv_cache.store(
+            key=key_to_cache,
+            value=value_to_cache,
+            slots=slots,
+            kv_scales=self.kv_scales,
+        )
+
+        # Prefill
+        if cu_seqlen_prefill is not None:
+            # flash attention
+            attn_output = attention(
+                query=query_states,
+                key=key_to_cache,
+                value=value_to_cache,
+                kv_cache=kv_cache,
+                kv_scales=self.kv_scales,
+                seqlen=seqlen,
+                block_tables=block_tables,
+                softmax_scale=self.softmax_scale,
+                window_size_left=self.window_size,
+            )
+        # Decode
+        else:
+            attn_output = paged_attention(
+                query_states,
+                kv_cache,
+                self.kv_head_mapping,
+                self.softmax_scale,
+                block_tables,
+                seqlen,
+                max_s,
+                kv_scales=self.kv_scales,
+                window_size_left=self.window_size,
+            )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        return self.o_proj(attn_output, adapter_data)
+
+
+# Import Qwen2MLP from the existing module to reuse the implementation
+from .flash_qwen2_modeling import Qwen2MLP as Qwen3MLP
+
+
+class Qwen3DecoderLayer(nn.Module):
+    def __init__(self, prefix, layer_id, config, weights):
+        super().__init__()
+        prefix = f"{prefix}.layers.{layer_id}"
+        self.hidden_size = config.hidden_size
+        self.self_attn = Qwen3Attention(
+            index=layer_id, prefix=f"{prefix}.self_attn", config=config, weights=weights
+        )
+        self.mlp = Qwen3MLP(
+            prefix=f"{prefix}.mlp", config=config, weights=weights, index=layer_id
+        )
+        self.input_layernorm = FastRMSNorm.load(
+            prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = FastRMSNorm.load(
+            prefix=f"{prefix}.post_attention_layernorm",
+            weights=weights,
+            eps=config.rms_norm_eps,
+        )
+
+    def forward(
+        self,
+        hidden_states,
+        residual,
+        cos,
+        sin,
+        cu_seqlen_prefill,
+        kv_cache,
+        block_tables,
+        slots,
+        seqlen,
+        max_s,
+        prefill_cache_indices,
+        adapter_data,
+    ):
+        residual = hidden_states
+        hidden_states, _ = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states = self.self_attn(
+            hidden_states,
+            cos,
+            sin,
+            cu_seqlen_prefill,
+            kv_cache,
+            block_tables,
+            slots,
+            seqlen,
+            max_s,
+            prefill_cache_indices,
+            adapter_data,
+        )
+
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states, _ = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states, adapter_data)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class Qwen3Model(torch.nn.Module):
+    def __init__(self, prefix: str, config, weights):
+        super().__init__()
+
+        prefix = f"{prefix}.model" if prefix else "model"
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        process_group = weights.process_group
+        self.tp_rank = process_group.rank()
+        self.tp_world_size = process_group.size()
+        self.layers = nn.ModuleList(
+            [
+                Qwen3DecoderLayer(
+                    prefix,
+                    layer_id,
+                    config,
+                    weights,
+                )
+                for layer_id in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = FastRMSNorm.load(
+            prefix=f"{prefix}.norm", weights=weights, eps=config.rms_norm_eps
+        )
+
+        self.gradient_checkpointing = False
+
+        self.head_size = self.layers[0].self_attn.head_dim
+        self.num_heads = self.layers[0].self_attn.num_heads
+        self.num_key_value_heads = self.layers[0].self_attn.num_key_value_heads
+
+    def forward(
+        self,
+        inputs_embeds: torch.Tensor,
+        position_ids: torch.Tensor,
+        cu_seqlen_prefill: Optional[torch.Tensor],
+        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
+        block_tables: torch.Tensor,
+        slots: torch.Tensor,
+        seqlen: Seqlen,
+        max_s: int,
+        true_max_s: int,
+        prefill_cache_indices: Optional[torch.Tensor],
+        adapter_data,
+    ) -> torch.Tensor:
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin(
+            position_ids,
+            true_max_s,
+            hidden_states.dtype,
+        )
+
+        residual = None
+        for i, decoder_layer in enumerate(self.layers):
+            hidden_states = decoder_layer(
+                hidden_states,
+                residual,
+                cos,
+                sin,
+                cu_seqlen_prefill,
+                kv_cache[i],
+                block_tables,
+                slots,
+                seqlen,
+                max_s,
+                prefill_cache_indices,
+                adapter_data,
+            )
+
+        hidden_states, _ = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        return hidden_states
+
+
+class Qwen3ForCausalLM(torch.nn.Module):
+    def __init__(self, prefix: str, config, weights):
+        super().__init__()
+
+        self.model = Qwen3Model(prefix, config, weights)
+        self.vocab_size = config.vocab_size
+
+        if config.tie_word_embeddings:
+            suffix = "model.embed_tokens"
+        else:
+            suffix = "lm_head"
+
+        self.lm_head = SpeculativeHead.load(
+            config,
+            prefix=f"{prefix}.{suffix}" if prefix else suffix,
+            weights=weights,
+        )
+
+        self.embed_tokens = TensorParallelEmbedding(
+            prefix=f"{prefix}.embed_tokens" if prefix else "model.embed_tokens",
+            weights=weights,
+        )
+
+        self.window_size = config.sliding_window
+        self.window_size_tensor = (
+            torch.tensor(config.sliding_window, device=weights.device)
+            if self.window_size is not None
+            else None
+        )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        cu_seqlen_prefill: Optional[torch.Tensor],
+        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
+        block_tables: torch.Tensor,
+        slots: torch.Tensor,
+        seqlen: Seqlen,
+        max_s: int,
+        prefill_cache_indices: Optional[torch.Tensor] = None,
+        lm_head_indices: Optional[torch.Tensor] = None,
+        adapter_data: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        true_max_s = max_s
+        if prefill_cache_indices is not None:
+            # Slots also need to be sliced as it has the same size as the whole kv tensor
+            slots = slots[prefill_cache_indices]
+        elif self.window_size is not None:
+            # Clamp in decode mode as paged attention requires clamped values whereas the flash attention
+            # kernel requires the true values
+            seqlen = seqlen.clamp(max=self.window_size_tensor)
+
+        inputs_embeds = self.embed_tokens(input_ids)
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        hidden_states = self.model(
+            inputs_embeds,
+            position_ids,
+            cu_seqlen_prefill,
+            kv_cache,
+            block_tables,
+            slots,
+            seqlen,
+            max_s,
+            true_max_s,
+            prefill_cache_indices,
+            adapter_data,
+        )
+
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        if lm_head_indices is not None:
+            hidden_states = hidden_states[lm_head_indices]
+        logits = self.lm_head(hidden_states)
+
+        return logits