feat: fix fp8 for MLA and support bmm fp8 for DeepSeek V2 (#1285)

Co-authored-by: ispobock <ispobaoke@163.com>

python/sglang/srt/models/deepseek_v2.py

@@ -19,6 +19,7 @@ limitations under the License.
 from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
+from flashinfer import bmm_fp8
 from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig
@@ -161,6 +162,15 @@ def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
     return 0.1 * mscale * math.log(scale) + 1.0
 
 
+def input_to_float8(x, dtype=torch.float8_e4m3fn):
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
+
+
 class DeepseekV2Attention(nn.Module):
 
     def __init__(
@@ -255,11 +265,6 @@ class DeepseekV2Attention(nn.Module):
             mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
             self.scaling = self.scaling * mscale * mscale
 
-        # self.attn = Attention(self.num_heads,
-        #                       self.qk_head_dim,
-        #                       self.scaling,
-        #                       num_kv_heads=self.num_heads)
-
         # TODO, support head_size 192
         self.attn = RadixAttention(
             self.num_local_heads,
@@ -283,7 +288,7 @@ class DeepseekV2Attention(nn.Module):
             q = self.q_proj(hidden_states)[0].view(
                 -1, self.num_local_heads, self.qk_head_dim
             )
-        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
         latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
         latent_cache = latent_cache.unsqueeze(1)
@@ -419,6 +424,7 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         self.w_kc = None
         self.w_vc = None
+        self.w_scale = None
 
     def forward(
         self,
@@ -439,8 +445,17 @@ class DeepseekV2AttentionMLA(nn.Module):
                 -1, self.num_local_heads, self.qk_head_dim
             )
         q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
-        q_nope_out = q_input[..., : self.kv_lora_rank]
-        torch.bmm(q_nope.transpose(0, 1), self.w_kc, out=q_nope_out.transpose(0, 1))
+
+        if self.w_kc.dtype == torch.float8_e4m3fn:
+            q_nope_val, q_nope_scale = input_to_float8(
+                q_nope.transpose(0, 1), torch.float8_e4m3fn
+            )
+            q_nope_out = bmm_fp8(
+                q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
+            )
+        else:
+            q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
+        q_input[..., : self.kv_lora_rank] = q_nope_out.transpose(0, 1)
 
         latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         v_input = latent_cache[..., : self.kv_lora_rank]
@@ -455,16 +470,21 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         attn_output = self.attn(q_input, k_input, v_input, input_metadata)
         attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
-        attn_bmm_output = attn_output.new_empty(
-            q_len, self.num_local_heads, self.v_head_dim
+
+        if self.w_vc.dtype == torch.float8_e4m3fn:
+            attn_output_val, attn_output_scale = input_to_float8(
+                attn_output.transpose(0, 1), torch.float8_e4m3fn
             )
-        torch.bmm(
-            attn_output.transpose(0, 1),
+            attn_bmm_output = bmm_fp8(
+                attn_output_val,
                 self.w_vc,
-            out=attn_bmm_output.transpose(0, 1),
+                attn_output_scale,
+                self.w_scale,
+                torch.bfloat16,
             )
-
-        attn_output = attn_bmm_output.flatten(1, 2)
+        else:
+            attn_bmm_output = torch.bmm(attn_output.transpose(0, 1), self.w_vc)
+        attn_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
         output, _ = self.o_proj(attn_output)
 
         return output
@@ -717,8 +737,10 @@ class DeepseekV2ForCausalLM(nn.Module):
                 w_kc, w_vc = self_attn.kv_b_proj.weight.unflatten(
                     0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
                 ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
-                self_attn.w_kc = w_kc.contiguous()
-                self_attn.w_vc = w_vc.transpose(1, 2).contiguous()
+                self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
+                self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
+                if hasattr(self_attn.kv_b_proj, "weight_scale"):
+                    self_attn.w_scale = self_attn.kv_b_proj.weight_scale
                 del self_attn.kv_b_proj