Correct Attention FLOPS estimation in flops_counter.py (#4929)

### What does this PR do?

Fix Attention FLOPS Calculation for Causal LLMs

### Problem

The current attention FLOPS calculations for all causal LLMs are missing
the `/2` factor for causal (lower triangular) attention mask. This
causes **2× overestimation** of attention FLOPS.

Additionally, DeepSeek V3's MLA attention incorrectly uses the same
dimension for both Q@K^T and attn@V operations, when `v_head_dim`
differs from `q_head_dim`.

### Changes Summary

| Function | Models | Change |
|----------|--------|--------|
| `_estimate_qwen2_flops` | qwen2, llama, qwen3, mistral, etc. | `12 *`
→ `6 *` |
| `_estimate_qwen3_vl_flops` | qwen3_vl | `12 *` → `6 *` |
| `_estimate_qwen3_vl_moe_flops` | qwen3_vl_moe | `12 *` → `6 *` |
| `_estimate_qwen2_moe_flops` | qwen2_moe, qwen3_moe | `12 *` → `6 *` |
| `_estimate_gemma3_flops` | gemma3_text | `12 *` → `6 *` |
| `_estimate_apertus_flops` | apertus | `12 *` → `6 *` |
| `_estimate_gpt_oss_flops` | gpt_oss | `12 *` → `6 *` |
| `_estimate_deepseek_v3_flops` | deepseek_v3 | `12 * q` → `3 * (q + v)`
|
| `_estimate_qwen3_vit_flop` | ViT (vision) | **No change**
(bidirectional) |

For causal (autoregressive) attention, only the lower triangular portion
of the attention matrix is computed:

```
Attention Matrix (causal):
[✓  ·  ·  ·]
[✓  ✓  ·  ·]
[✓  ✓  ✓  ·]
[✓  ✓  ✓  ✓]
```

| Model Type | Before | After | Overestimation |
|------------|--------|-------|----------------|
| Standard GQA/MHA | `12 * seq² * d` | `6 * seq² * d` | **2.0×** |
| DeepSeek V3 MLA | `12 * seq² * q` | `3 * seq² * (q+v)` | **2.4×** |

### Reference

This fix aligns with [Megatron-LM's FLOPS
calculation](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/training/training.py):
- Uses `/2` for causal attention
- Separately accounts for `q_head_dim` and `v_head_dim` in MLA

### Checklist Before Starting

- [ ] Search for similar PRs. Paste at least one query link here: ...
- [ ] Format the PR title as `[{modules}] {type}: {description}` (This
will be checked by the CI)
- `{modules}` include `fsdp`, `megatron`, `veomni`, `sglang`, `vllm`,
`rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`,
`deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`,
`model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward`
- If this PR involves multiple modules, separate them with `,` like
`[megatron, fsdp, doc]`
  - `{type}` is in `feat`, `fix`, `refactor`, `chore`, `test`
- If this PR breaks any API (CLI arguments, config, function signature,
etc.), add `[BREAKING]` to the beginning of the title.
  - Example: `[BREAKING][fsdp, megatron] feat: dynamic batching`

### Test

> For changes that can not be tested by CI (e.g., algorithm
implementation, new model support), validate by experiment(s) and show
results like training curve plots, evaluation results, etc.

### API and Usage Example

> Demonstrate how the API changes if any, and provide usage example(s)
if possible.

```python
# Add code snippet or script demonstrating how to use this
```

### Design & Code Changes

> Demonstrate the high-level design if this PR is complex, and list the
specific changes.

### Checklist Before Submitting

> [!IMPORTANT]
> Please check all the following items before requesting a review,
otherwise the reviewer might deprioritize this PR for review.

- [ ] Read the [Contribute
Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md).
- [ ] Apply [pre-commit
checks](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting):
`pre-commit install && pre-commit run --all-files --show-diff-on-failure
--color=always`
- [ ] Add / Update [the
documentation](https://github.com/volcengine/verl/tree/main/docs).
- [ ] Add unit or end-to-end test(s) to [the CI
workflow](https://github.com/volcengine/verl/tree/main/.github/workflows)
to cover all the code. If not feasible, explain why: ...
- [ ] Once your PR is ready for CI, send a message in [the `ci-request`
channel](https://verl-project.slack.com/archives/C091TCESWB1) in [the
`verl` Slack
workspace](https://join.slack.com/t/verl-project/shared_invite/zt-3855yhg8g-CTkqXu~hKojPCmo7k_yXTQ).
(If not accessible, please try [the Feishu group
(飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).)
- [ ] If your PR is related to the `recipe` submodule, please also
update the reference to the submodule commit via `git submodule update
--remote` or `cd recipe && git pull origin main`.

Author: HaochenYuan

verl/utils/flops_counter.py

@@ -112,7 +112,7 @@ def _estimate_qwen2_flops(config, tokens_sum, batch_seqlens, delta_time):
     seqlen_square_sum = 0
     for seqlen in batch_seqlens:
         seqlen_square_sum += seqlen * seqlen
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
 
     # all_layer & all_token fwd & bwd flops
     flops_all_token = dense_N_flops + attn_qkv_flops
@@ -149,7 +149,7 @@ def _estimate_qwen3_vl_flops(config, tokens_sum, batch_seqlens, delta_time, **ka
     seqlen_square_sum = 0
     for seqlen in batch_seqlens:
         seqlen_square_sum += seqlen * seqlen
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
 
     # vit flops
     images_seqlens = kargs.get("images_seqlens", None)
@@ -197,7 +197,7 @@ def _estimate_qwen3_vl_moe_flops(config, tokens_sum, batch_seqlens, delta_time,
     seqlen_square_sum = 0
     for seqlen in batch_seqlens:
         seqlen_square_sum += seqlen * seqlen
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
 
     # vit flops
     images_seqlens = kargs.get("images_seqlens", None)
@@ -304,7 +304,10 @@ def _estimate_deepseek_v3_flops(config, tokens_sum, batch_seqlens, delta_time):
     for seqlen in batch_seqlens:
         seqlen_square_sum += seqlen * seqlen * num_hidden_layers
 
-    attn_qkv_flops = 12 * seqlen_square_sum * q_head_dim * num_query_heads
+    # Core attention FLOPS for MLA with causal mask:
+    # Q @ K^T: 3 * 2 * seq^2 * q_head_dim * num_heads / 2 (causal)
+    # attn @ V: 3 * 2 * seq^2 * v_head_dim * num_heads / 2 (causal)
+    attn_qkv_flops = 3 * seqlen_square_sum * (q_head_dim + config.v_head_dim) * num_query_heads
     # all_layer & all_token fwd & bwk flops
     flops_all_token = dense_N_flops + attn_qkv_flops
     flops_achieved = flops_all_token * (1.0 / delta_time) / 1e12
@@ -341,7 +344,7 @@ def _estimate_qwen2_moe_flops(config, tokens_sum, batch_seqlens, delta_time):
     seqlen_square_sum = 0
     for seqlen in batch_seqlens:
         seqlen_square_sum += seqlen * seqlen
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
 
     # all_layer & all_token fwd & bwd flops
     flops_all_token = dense_N_flops + attn_qkv_flops
@@ -409,7 +412,7 @@ def _estimate_gemma3_flops(config, tokens_sum, batch_seqlens, delta_time):
             seqlen_square_sum += seqlen * seqlen
         seqlen_square_sum *= num_hidden_layers
 
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads
 
     # all_layer & all_token fwd & bwd flops
     flops_all_token = dense_N_flops + attn_qkv_flops
@@ -449,7 +452,7 @@ def _estimate_apertus_flops(config, tokens_sum, batch_seqlens, delta_time):
     seqlen_square_sum = 0
     for seqlen in batch_seqlens:
         seqlen_square_sum += seqlen * seqlen
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads * num_hidden_layers
 
     # all_layer & all_token fwd & bwd flops
     flops_all_token = dense_N_flops + attn_qkv_flops
@@ -520,7 +523,7 @@ def _estimate_gpt_oss_flops(config, tokens_sum, batch_seqlens, delta_time):
             seqlen_square_sum += seqlen * seqlen
         seqlen_square_sum *= num_hidden_layers
 
-    attn_qkv_flops = 12 * seqlen_square_sum * head_dim * num_attention_heads
+    attn_qkv_flops = 6 * seqlen_square_sum * head_dim * num_attention_heads
 
     # Total FLOPs
     flops_all_token = dense_N_flops + attn_qkv_flops