Enable renormalize(naive) routing for fp8 per-tensor (#2030)

<!-- .github/pull_request_template.md -->

## 📌 Description

Disable expert weights in the FC1 except for Llama routing.

## 🔍 Related Issues

<!-- Link any related issues here -->

## 🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull
request, please make sure the following items are complete.

### ✅ Pre-commit Checks

- [ ] I have installed `pre-commit` by running `pip install pre-commit`
(or used your preferred method).
- [ ] I have installed the hooks with `pre-commit install`.
- [ ] I have run the hooks manually with `pre-commit run --all-files`
and fixed any reported issues.

> If you are unsure about how to set up `pre-commit`, see [the
pre-commit documentation](https://pre-commit.com/).

## 🧪 Tests

- [ ] Tests have been added or updated as needed.
- [ ] All tests are passing (`unittest`, etc.).

## Reviewer Notes

<!-- Optional: anything you'd like reviewers to focus on, concerns, etc.
-->


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Bug Fixes**
  * Re-enabled Renormalize routing that was previously blocked.
  * Made token_scales available for Llama4 routing.
* Corrected GEMM1 input so the proper data source is used during MoE
processing.

* **Tests**
  * Added FP8PerTensorMoe to test parameterization.
* Expanded Renormalize and DeepSeekV3 test coverage and removed related
skips.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: siyuanf <[email protected]>

Author: IwakuraRein

csrc/trtllm_fused_moe_kernel_launcher.cu

@@ -584,6 +584,9 @@ class Fp8PerTensorLauncher : public FusedMoeLauncher {
         alloc_tensor({args->num_tokens, args->top_k}, dl_bfloat16, hidden_states.device());
 
     workspace.expert_weights = expert_weights.data_ptr();
+    if (static_cast<RoutingMethodType>(routing_method_type) == RoutingMethodType::Llama4) {
+      workspace.token_scales = expert_weights.data_ptr();  // Consumed by permuteGemm1 kernel
+    }
   }
 
   void check_moe() const override {

csrc/trtllm_fused_moe_runner.cu

@@ -518,7 +518,7 @@ void Runner::run(MoERunnerArgs const& args, MoEWorkspace const& workspace, int d
   auto const& config = mPassingConfigs[configIndex];
 
   mPermuteGemm1.run(args.hidden_states, hidden_states_scale_linear, args.gemm1_weights,
-                    args.gemm1_weights_scale, workspace.expert_weights, args.output1_scales_scalar,
+                    args.gemm1_weights_scale, workspace.token_scales, args.output1_scales_scalar,
                     args.output1_scales_gate_scalar, args.gemm1_bias, args.gemm1_alpha,
                     args.gemm1_beta, args.gemm1_clamp_limit, workspace.gemm1_output,
                     workspace.gemm1_output_scale, args.top_k, args.hidden_size,

include/flashinfer/trtllm/fused_moe/runner.h

@@ -305,7 +305,11 @@ struct MoEWorkspace {
   int32_t* expanded_idx_to_permuted_idx = nullptr;
   int32_t* permuted_idx_to_expanded_idx = nullptr;
   int32_t* permuted_idx_to_token_idx = nullptr;
+
+  // consumed by finalize kernel
   void* expert_weights = nullptr;  // [num_tokens, top_k] in bfloat16 = mDtypeExpW
+  // consumed by permuteGemm1 kernel
+  void* token_scales = nullptr;
 
   int32_t* cta_idx_xy_to_batch_idx = nullptr;
   int32_t* cta_idx_xy_to_mn_limit = nullptr;

tests/moe/test_trtllm_gen_fused_moe.py

@@ -2275,6 +2275,7 @@ def run_moe_test(
     [
         pytest.param(BF16Moe(), id="BF16xBF16"),
         pytest.param(FP8BlockScaleMoe(), id="FP8_Block"),
+        pytest.param(FP8PerTensorMoe(), id="FP8_Tensor"),
         pytest.param(FP4Moe(quant_mode=QuantMode.FP4_NVFP4_NVFP4), id="NvFP4xNvFP4"),
         pytest.param(FP4Moe(quant_mode=QuantMode.FP4_MXFP4_MXFP8), id="MxFP4xMxFP8"),
         pytest.param(FP4Moe(quant_mode=QuantMode.FP4_MXFP4_Bf16), id="MxFP4xBf16"),
@@ -2293,7 +2294,12 @@ def run_moe_test(
                 "routed_scaling": None,
                 "has_routing_bias": False,
                 "routing_method_type": RoutingMethodType.Renormalize,
-                "compatible_moe_impls": [FP8BlockScaleMoe, FP4Moe, BF16Moe],
+                "compatible_moe_impls": [
+                    FP8PerTensorMoe,
+                    FP8BlockScaleMoe,
+                    FP4Moe,
+                    BF16Moe,
+                ],
                 "compatible_intermediate_size": [384, 768, 1024],
             },
             id="Qwen3",
@@ -2308,7 +2314,12 @@ def run_moe_test(
                 "routed_scaling": None,
                 "has_routing_bias": False,
                 "routing_method_type": RoutingMethodType.Renormalize,
-                "compatible_moe_impls": [FP8BlockScaleMoe, FP4Moe, BF16Moe],
+                "compatible_moe_impls": [
+                    FP8PerTensorMoe,
+                    FP8BlockScaleMoe,
+                    FP4Moe,
+                    BF16Moe,
+                ],
                 "compatible_intermediate_size": [384, 1024],
             },
             id="Renorm",
@@ -2323,7 +2334,12 @@ def run_moe_test(
                 "routed_scaling": None,
                 "has_routing_bias": False,
                 "routing_method_type": RoutingMethodType.Renormalize,
-                "compatible_moe_impls": [FP8BlockScaleMoe, FP4Moe, BF16Moe],
+                "compatible_moe_impls": [
+                    FP8PerTensorMoe,
+                    FP8BlockScaleMoe,
+                    FP4Moe,
+                    BF16Moe,
+                ],
                 "compatible_intermediate_size": [512],
             },
             id="Qwen3_next",