[Gaudi] Enable Qwen3_moe model (#3244)

Signed-off-by: yuanwu <yuan.wu@intel.com>

backends/gaudi/server/text_generation_server/models/__init__.py

@@ -104,6 +104,9 @@ try:
     from text_generation_server.models.custom_modeling.flash_qwen3_modeling import (
         Qwen3ForCausalLM,
     )
+    from text_generation_server.models.custom_modeling.flash_qwen3_moe_modeling import (
+        Qwen3MoeForCausalLM,
+    )
     from text_generation_server.models.custom_modeling.flash_mistral_modeling import (
         FlashMistralForCausalLM,
     )
@@ -292,7 +295,11 @@ class ModelType(enum.Enum):
         "name": "Qwen 3",
         "url": "https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2e4f653967f",
     }
-
+    QWEN3_MOE = {
+        "type": "qwen3_moe",
+        "name": "Qwen 3 Moe",
+        "url": "https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2e4f653967f",
+    }
     GALACTICA = {
         "type": "galactica",
         "name": "Galactica",
@@ -808,6 +815,18 @@ def get_model(
                 trust_remote_code=trust_remote_code,
                 lora_adapter_ids=lora_adapter_ids,
             )
+        elif model_type == QWEN3_MOE:
+            return FlashCausalLM(
+                model_id=model_id,
+                model_class=Qwen3MoeForCausalLM,
+                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 model_type == MLLAMA:
             return FlashMllamaCausalLM(
                 model_id=model_id,

backends/gaudi/server/text_generation_server/models/custom_modeling/flash_qwen3_moe_modeling.py

@@ -0,0 +1,542 @@
+# coding=utf-8
+# Copyright 5 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Type
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from text_generation_server.layers.attention import (
+    attention,
+    paged_attention,
+    Seqlen,
+    HPUPagedAttentionMetadata,
+)
+from text_generation_server.layers.attention.kv_cache import get_kv_scales
+from text_generation_server.layers.moe import DenseMoELayer, MoELayer, SparseMoELayer
+from text_generation_server.layers import (
+    TensorParallelEmbedding,
+    TensorParallelColumnLinear,
+    TensorParallelRowLinear,
+    SpeculativeHead,
+    FastLinear,
+)
+
+from text_generation_server.layers.layernorm import (
+    FastRMSNorm,
+)
+from .flash_qwen2_modeling import Qwen2MLP
+from .flash_qwen3_modeling import Qwen3Attention
+from transformers.activations import ACT2FN
+from text_generation_server.layers.rotary import PositionRotaryEmbedding
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class Qwen3MoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config, prefix, weights, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(
+            config, "head_dim", config.hidden_size // config.num_attention_heads
+        )
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = (
+            config.num_attention_heads // config.num_key_value_heads
+        )
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = FastLinear.load(
+            config, f"{prefix}.q_proj", weights, bias=config.attention_bias
+        )
+
+        self.k_proj = FastLinear.load(
+            config, f"{prefix}.k_proj", weights, bias=config.attention_bias
+        )
+        self.v_proj = FastLinear.load(
+            config, f"{prefix}.v_proj", weights, bias=config.attention_bias
+        )
+        self.o_proj = FastLinear.load(
+            config, f"{prefix}.o_proj", weights, bias=config.attention_bias
+        )
+
+        self.rotary_emb = PositionRotaryEmbedding.static(
+            config=config,
+            dim=self.head_dim,
+            base=config.rope_theta,
+            device=weights.device,
+        )
+
+        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,
+        )
+
+        self.max_past = (
+            config.sliding_window if config.sliding_window is not None else -1
+        )
+
+        self.kv_scales = get_kv_scales(weights, f"{prefix}")
+        self.kv_head_mapping = torch.arange(
+            0, self.num_key_value_heads, dtype=torch.int32, device=weights.device
+        ).repeat_interleave(self.num_key_value_groups)
+
+        self.sliding_window = config.sliding_window
+        if not (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            self.sliding_window = None
+
+    def forward(
+        self,
+        hidden_states,
+        cos,
+        sin,
+        cu_seqlen_prefill,
+        kv_cache,
+        slots,
+        seqlen,
+        hpu_attention_meta,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states, _ = self.q_norm(self.q_proj(hidden_states).view(hidden_shape))
+        key_states, _ = self.k_norm(self.k_proj(hidden_states).view(hidden_shape))
+        value_states = self.v_proj(hidden_states).view(hidden_shape)
+
+        self.rotary_emb(query_states, key_states, cos, sin)
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        kv_cache.store(
+            key=key_states,
+            value=value_states,
+            slots=slots,
+            kv_scales=self.kv_scales,
+        )
+
+        # Prefill
+        if cu_seqlen_prefill is not None:
+            # sdpa
+            attn_output = attention(
+                query=query_states,
+                key=key_states,
+                value=value_states,
+                kv_cache=kv_cache,
+                kv_scales=self.kv_scales,
+                seqlen=seqlen,
+                softmax_scale=self.scaling,
+                window_size_left=self.max_past,
+            )
+        # Decode
+        else:
+            attn_output = paged_attention(
+                query_states,
+                kv_cache,
+                self.kv_head_mapping,
+                self.scaling,
+                seqlen,
+                kv_scales=self.kv_scales,
+                hpu_attention_meta=hpu_attention_meta,
+            )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output
+
+
+class Qwen3MoE(nn.Module):
+    def __init__(self, prefix, config, moe_layer_cls: Type[MoELayer], weights):
+        super().__init__()
+
+        # gating
+        self.gate = FastLinear.load(config, f"{prefix}.gate", weights, bias=False)
+
+        self.moe = moe_layer_cls(
+            n_expert_group=None,
+            n_experts=config.num_experts,
+            prefix=f"{prefix}.experts",
+            renormalize=True,
+            topk=config.num_experts_per_tok,
+            topk_group=None,
+            weights=weights,
+        )
+        # gate_proj_name="w1",
+        # up_proj_name="w3",
+        # down_proj_name="w2",
+
+        assert isinstance(self.moe, MoELayer)
+
+        self.process_group = weights.process_group
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        router_logits = self.gate(x)
+        out = self.moe(x, gating_output=router_logits)
+
+        # Reduce sum
+        if self.process_group.size() > 1:
+            torch.distributed.all_reduce(out, group=self.process_group)
+
+        return out.view(*x.shape)
+
+
+class Qwen3MoeMLP(nn.Module):
+    def __init__(self, prefix, config, weights, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = (
+            intermediate_size
+            if intermediate_size is not None
+            else config.intermediate_size
+        )
+        # Fuse gate and up proj
+        self.gate_up_proj = TensorParallelColumnLinear.load_multi(
+            config,
+            prefixes=[f"{prefix}.gate_proj", f"{prefix}.up_proj"],
+            weights=weights,
+            dim=0,
+            bias=False,
+        )
+        self.down_proj = TensorParallelRowLinear.load(
+            config,
+            prefix=f"{prefix}.down_proj",
+            weights=weights,
+            bias=False,
+        )
+        self.intermediate_size = (
+            config.intermediate_size // weights.process_group.size()
+        )
+
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        gate_up_states = self.gate_up_proj(x)
+        gate_up_states = gate_up_states.view(-1, 2, self.intermediate_size)
+        return self.down_proj(self.act(gate_up_states[:, 0]) * gate_up_states[:, 1])
+
+
+class Qwen3MoeSparseMoeBlock(nn.Module):
+    def __init__(self, prefix, config, weights):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+
+        # gating
+        # self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+        self.gate = FastLinear.load(config, f"{prefix}.gate", weights, bias=False)
+        self.experts = nn.ModuleList(
+            [
+                Qwen3MoeMLP(
+                    prefix=f"{prefix}.experts.{i}",
+                    config=config,
+                    weights=weights,
+                    intermediate_size=config.moe_intermediate_size,
+                )
+                for i in range(self.num_experts)
+            ]
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        input_shape = hidden_states.shape
+        _, hidden_dim = hidden_states.shape
+        # hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=hidden_states.dtype)
+        routing_weights, selected_experts = torch.topk(
+            routing_weights, self.top_k, dim=-1
+        )
+        if self.norm_topk_prob:  # only diff with mixtral sparse moe block!
+            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (input_shape), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(
+            selected_experts, num_classes=self.num_experts
+        ).permute(2, 1, 0)
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = (
+                expert_layer(current_state) * routing_weights[top_x, idx, None]
+            )
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(
+                0, top_x, current_hidden_states.to(hidden_states.dtype)
+            )
+        final_hidden_states = final_hidden_states.reshape(input_shape)
+        return final_hidden_states
+
+
+class Qwen3MoeDecoderLayer(nn.Module):
+    def __init__(self, config, prefix, weights, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        if config.num_key_value_heads // weights.process_group.size() > 0:
+            self.self_attn = Qwen3Attention(
+                config,
+                prefix=f"{prefix}.self_attn",
+                weights=weights,
+                layer_idx=layer_idx,
+            )
+        else:
+            self.self_attn = Qwen3MoeAttention(
+                config,
+                prefix=f"{prefix}.self_attn",
+                weights=weights,
+                layer_idx=layer_idx,
+            )
+
+        moe_layer_cls = (
+            SparseMoELayer if SparseMoELayer.is_supported(weights) else DenseMoELayer
+        )
+
+        if (layer_idx not in config.mlp_only_layers) and (
+            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        ):
+            self.mlp = Qwen3MoE(f"{prefix}.mlp", config, moe_layer_cls, weights)
+            # self.mlp = Qwen3MoeSparseMoeBlock(f"{prefix}.mlp", config, weights)
+
+        else:
+            self.mlp = Qwen2MLP(config=config, prefix=f"{prefix}.mlp", weights=weights)
+
+        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,
+        slots,
+        seqlen,
+        hpu_attention_meta,
+    ) -> torch.Tensor:
+        if residual is None:
+            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,
+            slots,
+            seqlen,
+            hpu_attention_meta,
+        )
+
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states, _ = self.post_attention_layernorm(hidden_states)
+
+        hidden_states = self.mlp(hidden_states)
+
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class Qwen3MoeModel(nn.Module):
+    def __init__(self, config, prefix: str, weights):
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.layers = nn.ModuleList(
+            [
+                Qwen3MoeDecoderLayer(
+                    config=config,
+                    prefix=f"{prefix}.layers.{layer_idx}",
+                    weights=weights,
+                    layer_idx=layer_idx,
+                )
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = FastRMSNorm.load(
+            prefix=f"{prefix}.norm", weights=weights, eps=config.rms_norm_eps
+        )
+
+    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]],
+        slots: torch.Tensor,
+        seqlen: Seqlen,
+        hpu_attention_meta: Optional[HPUPagedAttentionMetadata],
+    ) -> 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,
+        )
+
+        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],
+                slots,
+                seqlen,
+                hpu_attention_meta,
+            )
+
+        hidden_states, _ = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        return hidden_states
+
+
+class Qwen3MoeForCausalLM(nn.Module):
+
+    def __init__(self, prefix: str, config, weights):
+        super().__init__()
+        self.model = Qwen3MoeModel(config=config, prefix="model", weights=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,
+        )
+
+    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]],
+        slots: torch.Tensor,
+        seqlen: Seqlen,
+        hpu_attention_meta: Optional[HPUPagedAttentionMetadata],
+        lm_head_indices: Optional[torch.Tensor] = None,
+        adapter_data: Optional[torch.Tensor] = None,
+    ) -> torch.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,
+            slots,
+            seqlen,
+            hpu_attention_meta,
+        )
+
+        # 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

backends/gaudi/server/text_generation_server/models/flash_causal_lm.py

@@ -1469,7 +1469,7 @@ class FlashCausalLM(Model):
             raise ValueError("Cannot get the number of key/value heads")
         self.num_kv_heads = (
             num_kv_heads // self.process_group.size()
-            if num_kv_heads > 1
+            if num_kv_heads // self.process_group.size() > 0
             else num_kv_heads
         )
         assert self.num_kv_heads > 0