Mla deepspeek (#2)

* mla optimization

* hpu need padding in the first token generation


Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

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

@@ -12,6 +12,7 @@ from text_generation_server.layers.speculative import SpeculativeHead
 # Just to add the `load` methods.
 from text_generation_server.layers.layernorm import load_layer_norm
 from text_generation_server.layers.conv import load_conv2d
+from text_generation_server.layers.fp8 import Fp8Linear
 
 from text_generation_server.layers.lora import (
     LoraLinear,
@@ -27,6 +28,7 @@ __all__ = [
     "TensorParallelEmbedding",
     "SpeculativeHead",
     "LoraLinear",
+    "Fp8Linear",
     "TensorParallelMultiAdapterLinear",
     "TensorParallelAdapterRowLinear",
     "load_layer_norm",

backends/gaudi/server/text_generation_server/layers/attention/__init__.py

@@ -10,18 +10,21 @@ from .hpu import (
     SUPPORTS_WINDOWING,
     attention,
     paged_attention,
+    paged_attention_mla,
 )
 
 
 # KVCache needs `reshape_and_cache`, so ensure that it is defined already.
-from .kv_cache import KVCache, get_kv_scales
+from .kv_cache import KVCache, get_kv_scales, KVCompressCache
 
 __all__ = [
     "attention",
     "get_kv_scales",
     "paged_attention",
+    "paged_attention_mla",
     "SUPPORTS_WINDOWING",
     "KVCache",
+    "KVCompressCache",
     "Seqlen",
     "HPUPagedAttentionMetadata",
     "trim_seqlen_metadata",

backends/gaudi/server/text_generation_server/layers/attention/hpu.py

@@ -117,7 +117,7 @@ def paged_attention(
     hpu_attention_meta: HPUPagedAttentionMetadata,
 ):
     batch_size, head_num, head_size = query.shape
-    fp8_kv = kv_cache.key.dtype == torch.float8_e4m3fn
+    fp8_kv = kv_cache.dtype == torch.float8_e4m3fn
     output = ops.flat_pa(
         query=query.view(batch_size, 1, head_num * head_size),
         key_cache=kv_cache.key,
@@ -138,8 +138,39 @@ def paged_attention(
     return output.view(batch_size, head_num, head_size)
 
 
-__all__ = [
-    "SUPPORTS_WINDOWING",
-    "attention",
-    "paged_attention",
-]
+def paged_attention_mla(
+    query: torch.Tensor,
+    kv_cache: KVCache,
+    kv_head_mapping: torch.Tensor,
+    softmax_scale: float,
+    seqlen: Seqlen,
+    *,
+    kv_scales: KVScales,
+    softcap: Optional[float] = None,
+    hpu_attention_meta: HPUPagedAttentionMetadata,
+    kv_lora_rank: int = 0,
+):
+    batch_size, head_num, head_size = query.shape
+    fp8_kv = kv_cache.dtype == torch.float8_e4m3fn
+    output = ops.flat_pa_mla(
+        query=query,
+        key_cache=kv_cache.key,
+        value_cache=None,
+        block_list=hpu_attention_meta.block_list,
+        block_mapping=hpu_attention_meta.block_mapping,
+        block_bias=hpu_attention_meta.attn_bias,
+        block_groups=hpu_attention_meta.block_groups,
+        scale=softmax_scale,
+        matmul_qk_op=FP8Matmul(kv_scales.key_scale) if fp8_kv else Matmul(),
+        matmul_av_op=FP8Matmul(kv_scales.value_scale) if fp8_kv else Matmul(),
+        batch2block_matmul_op=Matmul(),
+        block2batch_matmul_op=Matmul(),
+        keys_fetch_func=FetchFromCache(1.0 / kv_scales.key_scale_cpu),
+        values_fetch_func=None,
+        kv_lora_rank=kv_lora_rank,
+    )
+    # Reshape the output tensor.
+    return output.view(batch_size, head_num, -1)
+
+
+__all__ = ["SUPPORTS_WINDOWING", "attention", "paged_attention", "paged_attention_mla"]

backends/gaudi/server/text_generation_server/layers/attention/kv_cache.py

@@ -108,6 +108,69 @@ class KVCache:
         )
 
 
+class KVCompressCache(KVCache):
+    """
+    Key-value cache for attention layers.
+    """
+
+    kv_cache: torch.Tensor
+
+    def __init__(
+        self,
+        *,
+        num_blocks: int,
+        head_size: int,
+        dtype: torch.dtype,
+        device: torch.device,
+    ):
+        """Construct the key-value cache for a layer."""
+        ## TODO FP8 kv cache support
+        if dtype is torch.float8_e5m2:
+            raise ValueError("torch.float8_e5m2 is not supported in hpu. ")
+
+        self.kv_cache = torch.zeros(
+            (num_blocks, BLOCK_SIZE, 1, head_size),
+            dtype=dtype,
+            device=device,
+        )
+
+    @property
+    def dtype(self):
+        """Get the data type of the cache."""
+        return self.kv_cache.dtype
+
+    @property
+    def key(self):
+        """Get the key cache."""
+
+        return self.kv_cache
+
+    @property
+    def value(self):
+        """Get the value cache."""
+
+        return self.kv_cache
+
+    def store(
+        self,
+        *,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        slots: torch.Tensor,
+        kv_scales: KVScales,
+    ):
+        """Store the key and value at the given slots."""
+        ## TODO FP8 kv cache support
+
+        block_idx = slots // BLOCK_SIZE
+        block_offset = slots % BLOCK_SIZE
+        if self.kv_cache.dtype == torch.float8_e4m3fn:
+            key = torch.ops.hpu.cast_to_fp8_v2(
+                key, kv_scales.key_scale, False, False, torch.float8_e4m3fn
+            )[0]
+        cache_ops.insert_or_update_cache(key, self.kv_cache, block_idx, block_offset)
+
+
 def paged_reshape_and_cache(
     key: torch.Tensor,
     value: torch.Tensor,

backends/gaudi/server/text_generation_server/models/custom_modeling/flash_deepseek_v3_modeling.py

@@ -28,11 +28,12 @@ from text_generation_server.layers import (
     TensorParallelEmbedding,
     TensorParallelRowLinear,
     get_linear,
+    Fp8Linear,
 )
 from text_generation_server.layers.attention import (
     Seqlen,
     attention,
-    paged_attention,
+    paged_attention_mla,
     HPUPagedAttentionMetadata,
 )
 from text_generation_server.layers.attention.kv_cache import KVCache, get_kv_scales
@@ -42,6 +43,18 @@ from text_generation_server.layers.rotary import PositionRotaryEmbedding, get_ms
 from text_generation_server.utils.weights import Weights
 
 
+def get_and_maybe_dequant_weights(layer: torch.nn.Module) -> torch.Tensor:
+    if isinstance(layer, Fp8Linear):
+        eye = torch.eye(
+            layer.qweight.shape[-1], dtype=torch.bfloat16, device=layer.qweight.device
+        )
+        dequant_weights = layer(eye)
+        del eye
+        # standardize to (output, input)
+        return dequant_weights.T
+    return layer.weight
+
+
 class DeepseekV3Config(PretrainedConfig):
     def __init__(
         self,
@@ -249,6 +262,44 @@ class DeepseekV3Attention(torch.nn.Module):
             0, self.num_key_value_heads, dtype=torch.int32, device=weights.device
         ).repeat_interleave(self.num_groups)
 
+        kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj.linear).T
+        kv_b_proj_weight = kv_b_proj_weight.view(
+            self.kv_lora_rank,
+            self.num_heads,
+            self.qk_nope_head_dim + self.value_head_size,
+        )
+        W_UK, W_UV = kv_b_proj_weight.split(
+            [self.qk_nope_head_dim, self.value_head_size], dim=-1
+        )
+        # Convert from (L, N, V) to (N, L, V)
+        self.W_UV = W_UV.transpose(0, 1)
+        # Convert from (L, N, P) to (N, P, L)
+        self.W_UK_T = W_UK.permute(1, 2, 0)
+
+    def _q_proj_and_k_up_proj(self, x):
+        q_proj = self.q_proj if self.q_lora_rank is None else self.q_b_proj
+        q_nope, q_pe = (
+            q_proj(x)
+            .view(-1, self.num_heads, self.head_size)
+            .split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+        )
+
+        # Convert from (B, N, P) to (N, B, P)
+        q_nope = q_nope.transpose(0, 1)
+        # Multiply (N, B, P) x (N, P, L) -> (N, B, L)
+        ql_nope = torch.bmm(q_nope, self.W_UK_T)
+        # Convert from (N, B, L) to (B, N, L)
+        return ql_nope.transpose(0, 1), q_pe
+
+    def _v_up_proj_and_o_proj(self, x):
+        # Convert from (B, N, L) to (N, B, L)
+        x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)
+        # Multiply (N, B, L) x (N, L, V) -> (N, B, V)
+        x = torch.bmm(x, self.W_UV)
+        # Convert from (N, B, V) to (B, N * V)
+        x = x.transpose(0, 1).reshape(-1, self.num_heads * self.value_head_size)
+        return self.o_proj(x)
+
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -261,14 +312,9 @@ class DeepseekV3Attention(torch.nn.Module):
         hpu_attention_meta: Optional[HPUPagedAttentionMetadata],
     ):
         if self.q_lora_rank is None:
-            query = self.q_proj(hidden_states)
+            hidden_states_or_q_c = hidden_states
         else:
-            query = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states))[0])
-        query = query.view(-1, self.num_heads, self.head_size)
-
-        _, query_pe = torch.split(
-            query, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
-        )
+            hidden_states_or_q_c = self.q_a_layernorm(self.q_a_proj(hidden_states))[0]
 
         compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
         compressed_kv, key_pe = torch.split(
@@ -276,13 +322,18 @@ class DeepseekV3Attention(torch.nn.Module):
         )
 
         key_pe = key_pe.view(-1, 1, self.qk_rope_head_dim)
-        kv = self.kv_b_proj(self.kv_a_layernorm(compressed_kv.contiguous())[0]).view(
-            -1, self.num_key_value_heads, self.qk_nope_head_dim + self.value_head_size
-        )
+        kv_c_normed = self.kv_a_layernorm(compressed_kv.contiguous())[0]
 
-        key_nope, value = torch.split(
-            kv, [self.qk_nope_head_dim, self.value_head_size], dim=-1
+        # Prefill
+        if cu_seqlen_prefill is not None:
+            q_proj = self.q_proj if self.q_lora_rank is None else self.q_b_proj
+            query = q_proj(hidden_states_or_q_c)
+            query = query.view(-1, self.num_heads, self.head_size)
+            query_nope, query_pe = torch.split(
+                query, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
             )
+        else:
+            query_nope, query_pe = self._q_proj_and_k_up_proj(hidden_states_or_q_c)
 
         batch_size, heads, head_dim = query_pe.shape
         query_pe = (
@@ -297,7 +348,30 @@ class DeepseekV3Attention(torch.nn.Module):
             .reshape(batch_size, heads, head_dim)
         )
         self.rotary_emb(query_pe, key_pe, cos, sin)
+        latent_vec_k = torch.concat(
+            (kv_c_normed, key_pe.view(-1, self.qk_rope_head_dim)), dim=-1
+        )
+        latent_vec_k = latent_vec_k.view(-1, self.qk_rope_head_dim + self.kv_lora_rank)
 
+        latent_vec_k = latent_vec_k.unflatten(0, (slots.size(0), -1))
+
+        kv_cache.store(
+            key=latent_vec_k,
+            value=None,
+            slots=slots,
+            kv_scales=self.kv_scales,
+        )
+
+        if cu_seqlen_prefill is not None:
+            kv = self.kv_b_proj(kv_c_normed).view(
+                -1,
+                self.num_key_value_heads,
+                self.qk_nope_head_dim + self.value_head_size,
+            )
+
+            key_nope, value = torch.split(
+                kv, [self.qk_nope_head_dim, self.value_head_size], dim=-1
+            )
             query[..., self.qk_nope_head_dim :] = query_pe
             key = torch.empty_like(query)
             key[..., : self.qk_nope_head_dim] = key_nope
@@ -315,15 +389,6 @@ class DeepseekV3Attention(torch.nn.Module):
                 value, (0, self.head_pad_size - self.value_head_size), value=0
             )
 
-        kv_cache.store(
-            key=key,
-            value=value,
-            slots=slots,
-            kv_scales=self.kv_scales,
-        )
-
-        # Prefill
-        if cu_seqlen_prefill is not None:
             # flash attention
             attn_output = attention(
                 query=query,
@@ -334,9 +399,15 @@ class DeepseekV3Attention(torch.nn.Module):
                 seqlen=seqlen,
                 softmax_scale=self.softmax_scale,
             )
-        # Decode
+            attn_output = attn_output[..., : self.value_head_size]
+
+            return self.o_proj(
+                attn_output.reshape(-1, self.num_heads * self.value_head_size)
+            )
         else:
-            attn_output = paged_attention(
+            # Decode
+            query = torch.cat([query_nope, query_pe], dim=-1)
+            attn_output = paged_attention_mla(
                 query,
                 kv_cache,
                 self.kv_head_mapping,
@@ -344,14 +415,10 @@ class DeepseekV3Attention(torch.nn.Module):
                 seqlen,
                 kv_scales=self.kv_scales,
                 hpu_attention_meta=hpu_attention_meta,
+                kv_lora_rank=self.kv_lora_rank,
             )
-
-        # Remove padding.
-        attn_output = attn_output[..., : self.value_head_size]
-
-        return self.o_proj(
-            attn_output.reshape(-1, self.num_heads * self.value_head_size)
-        )
+            attn_output = self._v_up_proj_and_o_proj(attn_output)
+            return attn_output
 
 
 class DeepseekV3MLP(nn.Module):

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

@@ -53,6 +53,7 @@ from text_generation_server.models.globals import (
 )
 from text_generation_server.layers.attention import (
     KVCache,
+    KVCompressCache,
     Seqlen,
     HPUPagedAttentionMetadata,
     trim_attn_metadata,
@@ -68,7 +69,9 @@ from text_generation_server.utils.import_utils import (
     synchronize,
     get_free_memory,
 )
-
+from text_generation_server.utils.prefill_chunking import (
+    get_max_prefill_tokens,
+)
 import vllm_hpu_extension.environment as environment
 import habana_frameworks.torch as htorch
 import itertools
@@ -1482,6 +1485,17 @@ class FlashCausalLM(Model):
     ):
         self.kv_cache = []
         empty_cache()
+        if self.config.model_type == "deepseek_v3":
+            self.kv_cache = [
+                KVCompressCache(
+                    num_blocks=num_blocks,
+                    head_size=self.config.kv_lora_rank + self.config.qk_rope_head_dim,
+                    dtype=dtype,
+                    device=device,
+                )
+                for _ in range(num_layers)
+            ]
+        else:
             self.kv_cache = [
                 KVCache(
                     num_blocks=num_blocks,
@@ -1511,8 +1525,14 @@ class FlashCausalLM(Model):
         # Inspired by the original implementation in [vllm](https://github.com/vllm-project/vllm)
         # Calculate the number of blocks that can be allocated with the free memory
         dtype_size = torch.tensor([], dtype=self.kv_cache_dtype).element_size()
+        if self.config.model_type == "deepseek_v3":
+            cache_block_size = BLOCK_SIZE * (
+                self.config.kv_lora_rank + self.config.qk_rope_head_dim
+            )
+        else:
             cache_block_size = BLOCK_SIZE * self.num_kv_heads * self.head_size
-        total_cache_size = self.num_layers * cache_block_size * 2 * dtype_size
+            cache_block_size = cache_block_size * 2
+        total_cache_size = self.num_layers * cache_block_size * dtype_size
 
         try:
             self.init_kv_cache(
@@ -1572,7 +1592,7 @@ class FlashCausalLM(Model):
             self.kv_cache_dtype,
             self.device,
         )
-        self.max_batch_prefill_tokens = max_input_tokens * len(batch)
+        self.max_batch_prefill_tokens = get_max_prefill_tokens()
         max_num_seqs = int(os.getenv("MAX_BATCH_SIZE"))
         HPUBucketingContext = get_bucketing_context()
         max_total_tokens_aligned = math.ceil(max_total_tokens / BLOCK_SIZE) * BLOCK_SIZE
@@ -1589,7 +1609,7 @@ class FlashCausalLM(Model):
         max_blocks = max(
             BLOCK_SIZE, max_num_seqs * max_total_tokens_aligned // BLOCK_SIZE
         )
-        self.bucketing_ctx.num_hpu_blocks = max_blocks
+        self.bucketing_ctx.num_hpu_blocks = min(max_blocks, num_blocks)
         if os.getenv("VLLM_SKIP_WARMUP", "false").lower() == "true":
             self.bucketing_ctx.generate_prompt_buckets()
             self.bucketing_ctx.generate_decode_buckets(
@@ -1616,6 +1636,8 @@ class FlashCausalLM(Model):
         for i, (batch_size, seq_len) in enumerate(
             reversed(self.bucketing_ctx.prompt_buckets)
         ):
+            if batch_size * seq_len > self.max_batch_prefill_tokens:
+                continue
             log_master(logger.info, f"warmup prefill seq {seq_len} bs {batch_size}")
             for index in range(warmup_times):
                 self.warmup_prefill(seq_len, batch_size, batch)

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

@@ -350,6 +350,8 @@ class FlashMllamaCausalLM(FlashVlmCausalLM):
         for i, (batch_size, seq_len) in enumerate(
             reversed(self.bucketing_ctx.prompt_buckets)
         ):
+            if batch_size * seq_len > self.max_batch_prefill_tokens:
+                continue
             log_master(logger.info, f"warmup prefill seq {seq_len} bs {batch_size}")
             for index in range(warmup_times):
                 self.warmup_prefill(seq_len, batch_size, batch)

backends/v3/src/queue.rs

@@ -8,6 +8,7 @@ use std::cmp::max;
 use std::collections::VecDeque;
 use text_generation_router::infer::InferError;
 use text_generation_router::infer::InferStreamResponse;
+use text_generation_router::usage_stats::Env;
 use text_generation_router::validation::{
     Chunk, ChunksToString, ValidGenerateRequest, ValidGrammar, ValidParameters,
     ValidStoppingParameters,
@@ -15,7 +16,6 @@ use text_generation_router::validation::{
 use tokio::sync::{mpsc, oneshot};
 use tokio::time::Instant;
 use tracing::{info_span, instrument, Instrument, Span};
-
 /// Queue entry
 #[derive(Debug)]
 pub(crate) struct Entry {
@@ -185,6 +185,9 @@ struct State {
 
     /// Paged Attention Block Allocation
     block_allocator: Option<BlockAllocator>,
+
+    /// indicate if it's hpu device, the hpu device needs padding to generate first token.
+    is_hpu_device: bool,
 }
 
 impl State {
@@ -214,6 +217,7 @@ impl State {
             speculate,
             support_chunking,
             block_allocator,
+            is_hpu_device: Env::new().is_hpu_device(),
         }
     }
 
@@ -368,6 +372,21 @@ impl State {
                         }
                     }
 
+                    //HPU padding for the prefill
+                    if self.is_hpu_device {
+                        max_input_length = max_input_length.max(entry.request.input_length);
+                        let actual_prefill_tokens_for_hpu =
+                            (batch.len() + 1) as u32 * max_input_length;
+
+                        if actual_prefill_tokens_for_hpu > prefill_token_budget {
+                            // Entry is over budget
+                            // Add it back to the front
+                            tracing::debug!("Over budget: prefill_tokens={actual_prefill_tokens_for_hpu} > {prefill_token_budget}");
+                            self.entries.push_front((id, entry));
+                            break 'entry_loop;
+                        }
+                    }
+
                     prefill_tokens += postfix_len;
 
                     Some(block_allocation)