[Deepseek V3.2] Change indexer weights_proj to fp32

hlu1 · hlu1 · commit 11c7304dcf31 · 2025-11-17T15:31:24.000-08:00
diff --git a/docs/basic_usage/deepseek_v32.md b/docs/basic_usage/deepseek_v32.md
@@ -129,6 +129,13 @@ Latency: 25.109 s
 Output throughput: 5226.235 token/s
 ```
 
+To test long-context accuracy, run gsm8k with `--num-shots 20`. The results are very close to the 8 shots results:
+```
+Accuracy: 0.956
+Invalid: 0.000
+Latency: 29.545 s
+Output throughput: 4418.617 token/s
+```
 
 ### Accuracy Test with `gpqa-diamond`
 
@@ -143,6 +150,65 @@ Repeat: 8, mean: 0.797
 Scores: ['0.808', '0.798', '0.808', '0.798', '0.783', '0.788', '0.803', '0.793']
 ```
 
+### Accuracy Test with `aime 2025`
+
+Prepare the environment by installing NeMo-Skills in the docker or your own virtual environment:
+
+```
+pip install git+https://github.com/NVIDIA/NeMo-Skills.git --ignore-installed blinker
+```
+
+Modify the [`jinja chat_template`](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/tokenizer_config.json#L34) by replacing
+
+```
+{% set thinking = false %}
+```
+with
+```
+{% set thinking = true %}
+```
+and save it to `chat_template_thinking.jinja`.
+
+Launch the SGLang server with the modified chat-template file:
+```
+python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention --chat-template chat_template_thinking.jinja
+```
+
+Run the following script to evaluate AIME 2025:
+```
+#! /bin/bash
+export NEMO_SKILLS_DISABLE_UNCOMMITTED_CHANGES_CHECK=1
+
+ns prepare_data aime25
+
+PORT=30000
+BACKEND=sglang
+MODEL="deepseek-ai/DeepSeek-V3.2-Exp"
+MODEL_NAME="dsv32-fp8"
+
+echo "Starting AIME25 evaluation with model $MODEL on port $PORT using backend $BACKEND..."
+ns eval \
+  --benchmarks=aime25:4 \
+  --server_type=$BACKEND \
+  --model=$MODEL \
+  --server_address=http://localhost:${PORT}/v1 \
+  --output_dir=nemo_skills_aime25_${MODEL_NAME}_output_${BACKEND}_$(date +%Y%m%d_%H%M%S) \
+  ++max_concurrent_requests=512 \
+  ++server.api_key=dummy \
+  ++inference.tokens_to_generate=64000
+```
+
+Test results:
+
+
+| evaluation_mode    | num_entries | avg_tokens | gen_seconds | symbolic_correct     | no_answer |
+|--------------------|-------------|------------|-------------|-----------------------|-----------|
+| pass@1[avg-of-4]   | 30          | 14410      | 1758        | 85.83% ± 4.19%        | 0.00%     |
+| majority@4         | 30          | 14410      | 1758        | 90.00%                | 0.00%     |
+| pass@4             | 30          | 14410      | 1758        | 93.33%                | 0.00%     |
+
+Note that the result of problem#3 with id `aime25-2` is marked as false by nemo-skills  but is actually correct because nemo-skills fails to match predicted_answer `016` with expected_answer `16`. If we add 1/30 = 3.33% to the results, the pass@1[avg-of-4] result matches with reference which is 89.3.
+
 
 ## DSA long sequence context parallel optimization(experimental)
 
diff --git a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
@@ -109,7 +109,6 @@ def __init__(
         prefix: str = "",
         quant_config: Optional[QuantizationConfig] = None,
         alt_stream: Optional[torch.cuda.Stream] = None,
-        fuse_wk_and_weights_proj: bool = False,
     ):
         super().__init__()
         self.hidden_size = hidden_size
@@ -120,7 +119,6 @@ def __init__(
         self.q_lora_rank = q_lora_rank
         self.layer_id = layer_id
         self.alt_stream = alt_stream
-        self.fuse_wk_and_weights_proj = fuse_wk_and_weights_proj
         self.nsa_enable_prefill_cp = is_nsa_enable_prefill_cp()
         if self.nsa_enable_prefill_cp:
             self.cp_size = get_attention_tp_size()
@@ -139,28 +137,22 @@ def __init__(
             quant_config=quant_config,
             prefix=add_prefix("wq_b", prefix),
         )
-        if self.fuse_wk_and_weights_proj:
-            self.fused_wk_and_weights_proj = ReplicatedLinear(
-                self.hidden_size,
-                self.head_dim + self.n_heads,
-                bias=False,
-                prefix=add_prefix("fused_wk_and_weights_proj", prefix),
-            )
-        else:
-            self.wk = ReplicatedLinear(
-                self.hidden_size,
-                self.head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("wk", prefix),
-            )
-            # NOTE: weight_proj is not quantized
-            self.weights_proj = ReplicatedLinear(
-                self.hidden_size,
-                self.n_heads,
-                bias=False,
-                prefix=add_prefix("weights_proj", prefix),
-            )
+
+        self.wk = ReplicatedLinear(
+            self.hidden_size,
+            self.head_dim,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("wk", prefix),
+        )
+        # NOTE: weights_proj in the checkpoint is stored in bf16, while the parameters here are stored in fp32 for convenience
+        self.weights_proj = ReplicatedLinear(
+            self.hidden_size,
+            self.n_heads,
+            bias=False,
+            params_dtype=torch.float32,
+            prefix=add_prefix("weights_proj", prefix),
+        )
         self.k_norm = LayerNorm(self.head_dim, dtype=torch.float32)
         self.rotary_emb = get_rope_wrapper(
             rope_head_dim,
@@ -189,7 +181,6 @@ def _get_q_k_bf16(
         enable_dual_stream: bool,
         forward_batch: ForwardBatch,
     ):
-        weights = None
         if enable_dual_stream:
             current_stream = torch.cuda.current_stream()
             self.alt_stream.wait_stream(current_stream)
@@ -206,12 +197,7 @@ def _get_q_k_bf16(
                 )
             with torch.cuda.stream(self.alt_stream):
                 # TODO we should also put DeepGEMM half SM here?
-                if self.fuse_wk_and_weights_proj:
-                    key, weights = self.fused_wk_and_weights_proj(x)[0].split(
-                        [self.head_dim, self.n_heads], dim=-1
-                    )
-                else:
-                    key, _ = self.wk(x)
+                key, _ = self.wk(x)
                 key = self.k_norm(key)
 
                 k_rope, _ = torch.split(
@@ -224,17 +210,10 @@ def _get_q_k_bf16(
         else:
             query, _ = self.wq_b(q_lora)
             query = rearrange(query, "l (h d) -> l h d", d=self.head_dim)
-
             q_rope, _ = torch.split(
                 query, [self.rope_head_dim, self.head_dim - self.rope_head_dim], dim=-1
             )
-
-            if self.fuse_wk_and_weights_proj:
-                key, weights = self.fused_wk_and_weights_proj(x)[0].split(
-                    [self.head_dim, self.n_heads], dim=-1
-                )
-            else:
-                key, _ = self.wk(x)
+            key, _ = self.wk(x)
             key = self.k_norm(key)
             k_rope, _ = torch.split(
                 key, [self.rope_head_dim, self.head_dim - self.rope_head_dim], dim=-1
@@ -266,21 +245,16 @@ def _get_q_k_bf16(
             query = rotate_activation(query)
             key = rotate_activation(key)
 
-        return query, key, weights
+        return query, key
 
     def _get_k_bf16(
         self,
         x: torch.Tensor,
         positions: torch.Tensor,
         enable_dual_stream: bool,
     ):
-        # Compute only key, skip query and weights (weights is discarded if fused)
-        if self.fuse_wk_and_weights_proj:
-            key, _ = self.fused_wk_and_weights_proj(x)[0].split(
-                [self.head_dim, self.n_heads], dim=-1
-            )
-        else:
-            key, _ = self.wk(x)
+        # Compute only key, skip query
+        key, _ = self.wk(x)
         key = self.k_norm(key)
         k_rope, _ = torch.split(
             key, [self.rope_head_dim, self.head_dim - self.rope_head_dim], dim=-1
@@ -779,7 +753,7 @@ def forward_cuda(
                 return_indices,
             )
 
-        query, key, weights = self._get_q_k_bf16(
+        query, key = self._get_q_k_bf16(
             q_lora, x, positions, enable_dual_stream, forward_batch=forward_batch
         )
 
@@ -808,8 +782,7 @@ def forward_cuda(
             index_k_scale=k_scale,
         )
 
-        if not self.fuse_wk_and_weights_proj:
-            weights, _ = self.weights_proj(x)
+        weights, _ = self.weights_proj(x.float())
         weights = self._get_logits_head_gate(weights, q_scale)
 
         if is_cuda():
@@ -1037,7 +1010,7 @@ def forward_npu(
         past_key_states = forward_batch.token_to_kv_pool.get_index_k_buffer(layer_id)
 
         x = x.view(-1, self.hidden_size)
-        weights = self.weights_proj(x)[0]
+        weights = self.weights_proj(x.float())[0]
         block_table = (
             block_table[: actual_seq_lengths_q.size()[0]] if is_prefill else block_table
         )
diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
@@ -238,17 +238,6 @@ def add_forward_absorb_core_attention_backend(backend_name):
         logger.info(f"Added {backend_name} to FORWARD_ABSORB_CORE_ATTENTION_BACKENDS.")
 
 
-def is_nsa_indexer_wk_and_weights_proj_fused(config, quant_config):
-    """
-    NSA Indexer wk and weights_proj can be fused in FP4 model because they are both in BF16
-    """
-    return (
-        is_deepseek_nsa(config)
-        and quant_config is not None
-        and quant_config.get_name() == "modelopt_fp4"
-    )
-
-
 class AttnForwardMethod(IntEnum):
     # Use multi-head attention
     MHA = auto()
@@ -1224,9 +1213,6 @@ def __init__(
                 quant_config=quant_config,
                 layer_id=layer_id,
                 alt_stream=alt_stream,
-                fuse_wk_and_weights_proj=is_nsa_indexer_wk_and_weights_proj_fused(
-                    config, quant_config
-                ),
             )
 
         self.kv_b_proj = ColumnParallelLinear(
@@ -3766,12 +3752,6 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=Fal
             self.config.q_lora_rank is not None
         )
         cached_a_proj = {} if fuse_qkv_a_proj else None
-        # Fuse wk and weights_proj when NSA Indexer is enabled and quant_config is FP4. For nextn, fp4 is disabled so we cannot fuse.
-        fuse_wk_and_weights_proj = (
-            is_nsa_indexer_wk_and_weights_proj_fused(self.config, self.quant_config)
-            and not is_nextn
-        )
-        cached_wk_and_weights_proj = {} if fuse_wk_and_weights_proj else None
 
         if is_nextn:
             nextn_layer_prefix = f"model.layers.{nextn_layer_id}"
@@ -3957,57 +3937,6 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=Fal
                                 )
                                 cached_a_proj.pop(q_a_proj_name)
                                 cached_a_proj.pop(kv_a_proj_name)
-                        elif fuse_wk_and_weights_proj and (
-                            "wk" in name or "weights_proj" in name
-                        ):
-                            cached_wk_and_weights_proj[name] = loaded_weight
-                            wk_name = (
-                                name
-                                if "wk" in name
-                                else name.replace("weights_proj", "wk")
-                            )
-                            weights_proj_name = (
-                                name
-                                if "weights_proj" in name
-                                else name.replace("wk", "weights_proj")
-                            )
-
-                            # When both wk and weights_proj has been cached, load the fused weight to parameter
-                            if (
-                                wk_name in cached_wk_and_weights_proj
-                                and weights_proj_name in cached_wk_and_weights_proj
-                            ):
-                                wk_weight = cached_wk_and_weights_proj[wk_name]
-                                weights_proj_weight = cached_wk_and_weights_proj[
-                                    weights_proj_name
-                                ]
-                                # todo dequantize wk for fp8
-                                assert wk_weight.dtype == weights_proj_weight.dtype
-                                fused_weight = torch.cat(
-                                    [wk_weight, weights_proj_weight], dim=0
-                                )
-                                param_name = (
-                                    name.replace("wk", "fused_wk_and_weights_proj")
-                                    if "wk" in name
-                                    else name.replace(
-                                        "weights_proj",
-                                        "fused_wk_and_weights_proj",
-                                    )
-                                )
-                                param = params_dict[param_name]
-
-                                weight_loader = getattr(
-                                    param, "weight_loader", default_weight_loader
-                                )
-                                maybe_executor_submit(
-                                    executor=executor,
-                                    futures=futures,
-                                    use_async=use_async_loading,
-                                    func=weight_loader,
-                                    func_args=(param, fused_weight),
-                                )
-                                cached_wk_and_weights_proj.pop(wk_name)
-                                cached_wk_and_weights_proj.pop(weights_proj_name)
                         else:
                             if (
                                 "k_scale" in name or "v_scale" in name
