Merge pull request #294 from kvcache-ai/feat-fast-MLA

Feat fast mla

ktransformers/models/custom_cache.py

@@ -51,13 +51,34 @@ class StaticCache(transformers.StaticCache):
         cache_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, self.head_dim)
         if config.architectures[0] == "DeepseekV2ForCausalLM" or config.architectures[0] == "DeepseekV3ForCausalLM":
             # TODO: for deepseek, cache_shape is different whether using Absorbed MLA, check it automatically
-            # key_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, config.qk_rope_head_dim + config.qk_nope_head_dim)
-            # value_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, config.v_head_dim)
-            key_shape = (max_batch_size, 1, self.max_cache_len, config.qk_rope_head_dim)
-            value_shape = (max_batch_size, 1, self.max_cache_len, config.kv_lora_rank)
+            self.page_size = 64
+            self.max_pages = (self.max_cache_len + self.page_size - 1) // self.page_size
+            latent_shape = (self.max_pages, self.page_size, 1, config.kv_lora_rank + config.qk_rope_head_dim)
+            self.kv_lora_rank = config.kv_lora_rank
+            self.qk_rope_head_dim = config.qk_rope_head_dim
+            # TODO: support real page table
+            self.page_table_map = dict()
+            self.page_table_list = []
+            for idx in range(config.num_hidden_layers):
+                if isinstance(device, dict):
+                    target_device = device[f"blk.{idx}.self_attn"]["generate_device"]
+                else:
+                    target_device = device
+                
+                if target_device not in self.page_table_map:
+                    page_table = torch.zeros((max_batch_size, self.max_pages), dtype=torch.int32, device=target_device)
+                    for seq_id in range(max_batch_size):
+                        page_table[seq_id, :] = torch.arange(seq_id * self.max_pages, seq_id * self.max_pages + self.max_pages, dtype=torch.int32, device=target_device)
+                    self.page_table_map[target_device] = page_table
+                    
+                self.page_table_list.append(self.page_table_map[target_device])
+                    
+            self.is_MLA = True
+            self.is_page = True
         else:
             key_shape = cache_shape
             value_shape = cache_shape
+            self.is_MLA = False
 
         self.past_tokens = []
         self.num_hidden_layers = config.num_hidden_layers
@@ -68,10 +89,17 @@ class StaticCache(transformers.StaticCache):
                 target_device = device[f"blk.{idx}.self_attn"]["generate_device"]
             else:
                 target_device = device
-            new_layer_key_cache = torch.zeros(key_shape, dtype=self.dtype, device=target_device)
-            new_layer_value_cache = torch.zeros(value_shape, dtype=self.dtype, device=target_device)
-            torch._dynamo.mark_static_address(new_layer_key_cache)
-            torch._dynamo.mark_static_address(new_layer_value_cache)
+            
+            if self.is_MLA:
+                new_layer_key_cache = torch.zeros(latent_shape, dtype=self.dtype, device=target_device)
+                new_layer_value_cache = None
+                torch._dynamo.mark_static_address(new_layer_key_cache)
+            else:
+                new_layer_key_cache = torch.zeros(key_shape, dtype=self.dtype, device=target_device)
+                new_layer_value_cache = torch.zeros(value_shape, dtype=self.dtype, device=target_device)
+                torch._dynamo.mark_static_address(new_layer_key_cache)
+                torch._dynamo.mark_static_address(new_layer_value_cache)
+                
             self.key_cache.append(new_layer_key_cache)
             self.value_cache.append(new_layer_value_cache)
             self.past_tokens.append(0)
@@ -104,11 +132,21 @@ class StaticCache(transformers.StaticCache):
         cache_position = cache_kwargs.get("cache_position")
         k_out = self.key_cache[layer_idx]
         v_out = self.value_cache[layer_idx]
-        #print(cache_position)
-        k_out[:, :, cache_position] = key_states
-        v_out[:, :, cache_position] = value_states
         self.past_tokens[layer_idx] += cache_position.size(0)
-        return k_out, v_out
+        #print(cache_position)
+        if self.is_MLA:
+            page_idx = cache_position // self.page_size
+            page_offset = cache_position % self.page_size
+            # key shape (self.max_pages, self.page_size, 1, config.kv_lora_rank + config.qk_rope_head_dim)
+            #print("page_idx", page_idx)
+            #print("page_offset", page_offset)
+            k_out[page_idx, page_offset, :, :self.kv_lora_rank] = key_states
+            k_out[page_idx, page_offset, :, self.kv_lora_rank:] = value_states
+            return k_out, self.page_table_list[layer_idx]
+        else:
+            k_out[:, :, cache_position] = key_states
+            v_out[:, :, cache_position] = value_states
+            return k_out, v_out
 
     def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
         """Returns the sequence length of the cached states that were seen by the model."""

ktransformers/operators/triton_attention.py

+import triton
+import triton.language as tl
+
+@triton.jit
+def tanh(x):
+    # Tanh is just a scaled sigmoid
+    return 2 * tl.sigmoid(2 * x) - 1
+
+@triton.jit
+def _fwd_grouped_kernel_stage1(
+    Q,
+    K_Buffer,
+    V_Buffer,
+    sm_scale,
+    Req_to_tokens,
+    B_Seqlen,
+    Att_Out,
+    stride_req_to_tokens_b,
+    stride_qbs,
+    stride_qh,
+    stride_buf_kbs,
+    stride_buf_kh,
+    stride_buf_vbs,
+    stride_buf_vh,
+    stride_mid_ob,
+    stride_mid_oh,
+    stride_mid_os,
+    kv_group_num: tl.constexpr,
+    q_head_num: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_H: tl.constexpr,
+    NUM_KV_SPLITS: tl.constexpr,
+    PAGE_SIZE: tl.constexpr,
+    logit_cap: tl.constexpr,
+    Lk: tl.constexpr,
+    Lv: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_head_id = tl.program_id(1)
+    cur_kv_head = cur_head_id // tl.cdiv(kv_group_num, BLOCK_H)
+    split_kv_id = tl.program_id(2)
+
+    if kv_group_num > BLOCK_H:
+        VALID_BLOCK_H: tl.constexpr = BLOCK_H
+    else:
+        VALID_BLOCK_H: tl.constexpr = kv_group_num
+    cur_head = cur_head_id * VALID_BLOCK_H + tl.arange(0, BLOCK_H)
+    mask_h = cur_head < (cur_head_id + 1) * VALID_BLOCK_H
+    mask_h = mask_h & (cur_head < q_head_num)
+
+    offs_d = tl.arange(0, BLOCK_DMODEL)
+    offs_dv = tl.arange(0, BLOCK_DV)
+    mask_d = offs_d < Lk
+    mask_dv = offs_dv < Lv
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+    cur_batch_req_idx = cur_batch
+
+    offs_q = cur_batch * stride_qbs + cur_head[:, None] * stride_qh + offs_d[
+        None, :]
+    q = tl.load(Q + offs_q,
+                mask=(mask_h[:, None]) & (mask_d[None, :]),
+                other=0.0)
+
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        mask_dpe = offs_dpe < Lk
+        off_qpe = (cur_batch * stride_qbs + cur_head[:, None] * stride_qh +
+                   offs_dpe[None, :])
+        qpe = tl.load(Q + off_qpe,
+                      mask=(mask_h[:, None]) & (mask_dpe[None, :]),
+                      other=0.0)
+
+    kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
+    split_kv_start = kv_len_per_split * split_kv_id
+    split_kv_end = tl.minimum(split_kv_start + kv_len_per_split,
+                              cur_batch_seq_len)
+    
+    e_max = tl.zeros([BLOCK_H], dtype=tl.float32) - float("inf")
+    e_sum = tl.zeros([BLOCK_H], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_H, BLOCK_DV], dtype=tl.float32)
+
+    if split_kv_end > split_kv_start:
+        for start_n in range(split_kv_start, split_kv_end, BLOCK_N):
+            offs_n = start_n + tl.arange(0, BLOCK_N)
+            kv_page_number = tl.load(
+                Req_to_tokens + stride_req_to_tokens_b * cur_batch_req_idx +
+                offs_n // PAGE_SIZE,
+                mask=offs_n < split_kv_end,
+                other=0,
+            )
+            kv_loc = kv_page_number * PAGE_SIZE + offs_n % PAGE_SIZE
+            offs_buf_k = (kv_loc[None, :] * stride_buf_kbs +
+                          cur_kv_head * stride_buf_kh + offs_d[:, None])
+            k = tl.load(
+                K_Buffer + offs_buf_k,
+                mask=(offs_n[None, :] < split_kv_end) & (mask_d[:, None]),
+                other=0.0,
+            )
+            qk = tl.dot(q, k.to(q.dtype))
+            
+            if BLOCK_DPE > 0:
+                offs_buf_kpe = (kv_loc[None, :] * stride_buf_kbs +
+                                cur_kv_head * stride_buf_kh +
+                                offs_dpe[:, None])
+                kpe = tl.load(
+                    K_Buffer + offs_buf_kpe,
+                    mask=(offs_n[None, :] < split_kv_end) &
+                    (mask_dpe[:, None]),
+                    other=0.0,
+                )
+                qk += tl.dot(qpe, kpe.to(qpe.dtype))
+            qk *= sm_scale
+
+            if logit_cap > 0:
+                qk = logit_cap * tanh(qk / logit_cap)
+
+            qk = tl.where(mask_h[:, None] & (offs_n[None, :] < split_kv_end),
+                          qk, float("-inf"))
+
+            offs_buf_v = (kv_loc[:, None] * stride_buf_vbs +
+                          cur_kv_head * stride_buf_vh + offs_dv[None, :])
+            v = tl.load(
+                V_Buffer + offs_buf_v,
+                mask=(offs_n[:, None] < split_kv_end) & (mask_dv[None, :]),
+                other=0.0,
+            )
+
+            n_e_max = tl.maximum(tl.max(qk, 1), e_max)
+            re_scale = tl.exp(e_max - n_e_max)
+            p = tl.exp(qk - n_e_max[:, None])
+            acc *= re_scale[:, None]
+            acc += tl.dot(p.to(v.dtype), v)
+
+            e_sum = e_sum * re_scale + tl.sum(p, 1)
+            e_max = n_e_max
+
+        offs_mid_o = (cur_batch * stride_mid_ob +
+                      cur_head[:, None] * stride_mid_oh +
+                      split_kv_id * stride_mid_os + offs_dv[None, :])
+
+        tl.store(
+            Att_Out + offs_mid_o,
+            acc / e_sum[:, None],
+            mask=(mask_h[:, None]) & (mask_dv[None, :]),
+        )
+
+        offs_mid_o_1 = (cur_batch * stride_mid_ob + cur_head * stride_mid_oh +
+                        split_kv_id * stride_mid_os + Lv)
+
+        tl.store(
+            Att_Out + offs_mid_o_1,
+            e_max + tl.log(e_sum),
+            mask=mask_h,
+        )
+
+def _decode_grouped_att_m_fwd(
+    q,
+    k_buffer,
+    v_buffer,
+    att_out,
+    Req_to_tokens,
+    B_Seqlen,
+    num_kv_splits,
+    sm_scale,
+    page_size,
+    logit_cap,
+):
+    BLOCK = 32
+    Lk = k_buffer.shape[-1]
+    Lv = v_buffer.shape[-1]
+
+    # [TODO] work around shmem limit on MI3xx
+    
+    # TODO: support hip
+    #if is_hip_ and Lk >= 576:
+    #    BLOCK = 16
+
+    if Lk == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    elif Lk == 288:
+        BLOCK_DMODEL = 256
+        BLOCK_DPE = 32
+    else:
+        BLOCK_DMODEL = triton.next_power_of_2(Lk)
+        BLOCK_DPE = 0
+    BLOCK_DV = triton.next_power_of_2(Lv)
+
+    batch, head_num = q.shape[0], q.shape[1]
+    kv_group_num = q.shape[1] // k_buffer.shape[-2]
+
+    BLOCK_H = 16
+    NUM_KV_SPLITS = num_kv_splits
+    grid = (
+        batch,
+        triton.cdiv(head_num, min(BLOCK_H, kv_group_num)),
+        NUM_KV_SPLITS,
+    )
+
+    extra_kargs = {}
+    # TODO: support hip
+    """
+    if is_hip_:
+        # https://rocm.docs.amd.com/en/docs-6.2.0/how-to/llm-fine-tuning-optimization/optimizing-triton-kernel.html
+        # https://github.com/triton-lang/triton/blob/main/third_party/amd/backend/compiler.py
+        extra_kargs = {
+            "waves_per_eu": 4,
+            "matrix_instr_nonkdim": 16,
+            "kpack": 2
+        }
+    """
+    
+    _fwd_grouped_kernel_stage1[grid](
+        q,
+        k_buffer,
+        v_buffer,
+        sm_scale,
+        Req_to_tokens,
+        B_Seqlen,
+        att_out,
+        Req_to_tokens.stride(0),
+        q.stride(0),
+        q.stride(1),
+        k_buffer.stride(-3),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        k_buffer.stride(-2),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        v_buffer.stride(-3),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        v_buffer.stride(-2),  # Assume (..., PAGE_SIZE, NUM_HEADS, HEAD_DIM)
+        att_out.stride(0),
+        att_out.stride(1),
+        att_out.stride(2),
+        kv_group_num=kv_group_num,
+        q_head_num=head_num,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
+        BLOCK_DV=BLOCK_DV,
+        BLOCK_N=BLOCK,
+        BLOCK_H=BLOCK_H,
+        NUM_KV_SPLITS=NUM_KV_SPLITS,
+        PAGE_SIZE=page_size,
+        logit_cap=logit_cap,
+        num_warps=4,
+        num_stages=2,
+        Lk=Lk,
+        Lv=Lv,
+        **extra_kargs,
+    )
+
+@triton.jit
+def _fwd_kernel_stage2(
+    Mid_O,
+    o,
+    B_Seqlen,
+    stride_mid_ob,
+    stride_mid_oh,
+    stride_mid_os,
+    stride_obs,
+    stride_oh,
+    NUM_KV_SPLITS: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
+    Lv: tl.constexpr,
+):
+    cur_batch = tl.program_id(0)
+    cur_head = tl.program_id(1)
+
+    cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+
+    offs_d = tl.arange(0, BLOCK_DV)
+    mask_d = offs_d < Lv
+
+    e_sum = 0.0
+    e_max = -float("inf")
+    acc = tl.zeros([BLOCK_DV], dtype=tl.float32)
+
+    offs_v = cur_batch * stride_mid_ob + cur_head * stride_mid_oh + offs_d
+    offs_logic = cur_batch * stride_mid_ob + cur_head * stride_mid_oh + Lv
+
+    for split_kv_id in range(0, NUM_KV_SPLITS):
+        kv_len_per_split = tl.cdiv(cur_batch_seq_len, NUM_KV_SPLITS)
+        split_kv_start = kv_len_per_split * split_kv_id
+        split_kv_end = tl.minimum(split_kv_start + kv_len_per_split,
+                                  cur_batch_seq_len)
+
+        if split_kv_end > split_kv_start:
+            tv = tl.load(Mid_O + offs_v + split_kv_id * stride_mid_os,
+                         mask=mask_d,
+                         other=0.0)
+            tlogic = tl.load(Mid_O + offs_logic + split_kv_id * stride_mid_os)
+            n_e_max = tl.maximum(tlogic, e_max)
+
+            old_scale = tl.exp(e_max - n_e_max)
+            acc *= old_scale
+            exp_logic = tl.exp(tlogic - n_e_max)
+            acc += exp_logic * tv
+
+            e_sum = e_sum * old_scale + exp_logic
+            e_max = n_e_max
+
+    tl.store(
+        o + cur_batch * stride_obs + cur_head * stride_oh + offs_d,
+        acc / e_sum,
+        mask=mask_d,
+    )
+
+def _decode_softmax_reducev_fwd(
+    logits,
+    q,
+    o,
+    v_buffer,
+    b_seq_len,
+    num_kv_splits,
+):
+    batch, head_num = q.shape[0], q.shape[1]
+    Lv = v_buffer.shape[-1]
+    BLOCK_DV = triton.next_power_of_2(Lv)
+
+    NUM_KV_SPLITS = num_kv_splits
+
+    extra_kargs = {}
+    # TODO: support hip
+    """
+    if is_hip_:
+        # https://rocm.docs.amd.com/en/docs-6.2.0/how-to/llm-fine-tuning-optimization/optimizing-triton-kernel.html
+        # https://github.com/triton-lang/triton/blob/main/third_party/amd/backend/compiler.py
+        extra_kargs = {
+            "waves_per_eu": 4,
+            "matrix_instr_nonkdim": 16,
+            "kpack": 2
+        }
+    """
+    
+    grid = (batch, head_num)
+    _fwd_kernel_stage2[grid](
+        logits,
+        o,
+        b_seq_len,
+        logits.stride(0),
+        logits.stride(1),
+        logits.stride(2),
+        o.stride(0),
+        o.stride(1),
+        NUM_KV_SPLITS=NUM_KV_SPLITS,
+        BLOCK_DV=BLOCK_DV,
+        Lv=Lv,
+        num_warps=4,
+        num_stages=2,
+        **extra_kargs,
+    )
+
+def decode_attention_fwd_grouped(
+    q,
+    k_buffer,
+    v_buffer,
+    o,
+    req_to_token,
+    b_seq_len,
+    attn_logits,
+    num_kv_splits,
+    sm_scale,
+    page_size,
+    logit_cap=0.0,
+):
+    _decode_grouped_att_m_fwd(
+        q,
+        k_buffer,
+        v_buffer,
+        attn_logits,
+        req_to_token,
+        b_seq_len,
+        num_kv_splits,
+        sm_scale,
+        page_size,
+        logit_cap,
+    )
+
+    _decode_softmax_reducev_fwd(attn_logits, q, o, v_buffer, b_seq_len,
+                                num_kv_splits)
\ No newline at end of file

ktransformers/util/utils.py

@@ -133,7 +133,7 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
             )
         else:
             past_key_values = None
-        cache_position = torch.arange(seq_length, device=torch_device)
+        cache_position = torch.arange(seq_length, device=torch_device, dtype=torch.long)
         generated_ids = torch.zeros(
             batch_size, seq_length + max_new_tokens + 1, dtype=torch.int, device=torch_device
         )
@@ -178,7 +178,7 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
         generated_ids[:, seq_length] = next_token
         tokens.append(int(next_token))
         inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
-        cache_position = torch.tensor([seq_length], device=torch_device)
+        cache_position = torch.tensor([seq_length], device=torch_device, dtype=torch.long)
         position_ids = cache_position.unsqueeze(0)
         seq_length += 1