[Misc][Quantization] Remove kvcache quant weight related patches.

vllm_ascend/patch/__init__.py

@@ -829,44 +829,7 @@
 #    Future Plan:
 #       Remove this patch when:
 #       vLLM itself supports kv transfer for mamba
-# ** 21. File: worker/patch_weight_utils.py**
-# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-#   1. `vllm.model_executor.models.deepseek_v2.DeepseekV2ForCausalLM.load_weights`
-#    Why:
-#       The C8 weight quantized by modelslim will modify the model structure,
-#       and the scale and offset required for kvcache quantization will increase.
-#       In addition, the names of the quantization parameters are different from
-#       those in the community.
-#    How：
-#       we have enhanced the maybe_remap_kv_scale_name function.
-#    Future Plan:
-#       The maybe_remap_kv_scale_name function of the community is reconstructed to support
-#       multiple backends.
-# ** 21b. File: worker/patch_process_weights_after_loading.py**
-# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-#   1. `vllm.model_executor.model_loader.utils.process_weights_after_loading`
-#      `vllm.model_executor.model_loader.base_loader.process_weights_after_loading`
-#      and imported references in vllm-ascend model loaders
-#    Why:
-#       DSA attention is implemented in vllm-ascend as the plugin layer
-#       `DSAAttention`. Upstream vLLM only runs post-load attention weight
-#       processing for built-in attention classes, so
-#       `DSAAttention.process_weights_after_loading()` is skipped in the
-#       original loader flow. DSV4 DSA-CP o-proj TP initialization must run in
-#       this post-load phase rather than being initialized lazily in forward.
-#    How:
-#       Rebind the upstream `process_weights_after_loading` helper, including
-#       already-imported loader references, so `DSAAttention` participates in
-#       the same post-load traversal while preserving the original quant-method
-#       and torchao reload behavior.
-#    Related PR (if no, explain why):
-#       https://github.com/vllm-project/vllm-ascend/pull/10694
-#       https://github.com/vllm-project/vllm/pull/46828
-#    Future Plan:
-#       Remove this patch once the supported vLLM version includes PR #46828.
-#       Then register `DSAAttention` through vLLM's post-load weight-processing
-#       registry instead of monkey-patching model-loader helpers.
-# ** 22. File: worker/patch_v2/patch_input_batch.py**
+# ** 21. File: worker/patch_v2/patch_input_batch.py
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.worker.gpu.input_batch.InputBatch`
 #    Why:
@@ -876,7 +839,7 @@
 #       replace InputBatch with AscendInputBatch.
 #    Future Plan:
 #       remove this patch when vLLM-ascend's make_dummy behavior aligns with vLLM.
-# ** 23. File: worker/patch_v2/patch_block_table.py**
+# ** 22. File: worker/patch_v2/patch_block_table.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.worker.gpu.block_table.BlockTables`
 #    Why:
@@ -888,7 +851,7 @@
 #    Future Plan:
 #       remove this patch when vLLM-ascend's BlockTables can initialize
 #       slot mapping as torch.int64 dtype.
-# ** 24. File: worker/patch_v2/patch_model_state.py**
+# ** 23. File: worker/patch_v2/patch_model_state.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.worker.gpu.model_states.default.init_model_state`
 #    Why:
@@ -898,7 +861,7 @@
 #       Define AscendModelState and initialize it in init_model_state.
 #    Future Plan:
 #       remove this when vllm-ascend's attention metadata is align with vllm.
-# ** 25. File: worker/patch_v2/patch_triton.py**
+# ** 24. File: worker/patch_v2/patch_triton.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.worker.gpu.sample.logprob`, `vllm.v1.worker.gpu.sample.penalties.apply_penalties`,
 #      `vllm.v1.worker.gpu.sample.gumbel.gumbel_sample`
@@ -911,30 +874,7 @@
 #    Future Plan:
 #       Remove this patch when vLLM support the dispatch function.
 #
-# ** 26. File: worker/patch_gqa_c8.py**
-# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-#   1. `vllm.model_executor.models.qwen3.Qwen3ForCausalLM.load_weights`
-#    Why:
-#       The GQA W8A8C8 model stores per-channel KV cache scales and offsets
-#       (k_cache_scale, k_cache_offset, v_cache_scale, v_cache_offset) under
-#       weight names that AutoWeightsLoader does not recognise and would
-#       silently discard.  Without these scales the INT8 KV cache cannot be
-#       dequantised correctly at inference time.
-#    How:
-#       Wrap load_weights to intercept the C8 scale/offset tensors before they
-#       reach the base loader.  Each intercepted tensor is routed to the
-#       corresponding nn.Parameter via its weight_loader, then excluded from
-#       the remaining weight stream so the base loader never sees it.
-#    Related PR (if no, explain why):
-#       This PR (Qwen3-32B and GLM4.7  W8A8C8 support).  Upstream vLLM's weight-loading
-#       pipeline does not yet have a generic hook for hardware-plugin-defined
-#       KV cache parameters.
-#    Future Plan:
-#       Remove this patch when vLLM provides a first-class extension point
-#       for loading extra KV cache quantisation parameters in model load_weights,
-#       or when the GQA model's weight names are aligned with the parameter
-#       names expected by the quantisation backend.
-# ** 27. File: worker/patch_qwen3vl.py**
+# ** 25. File: worker/patch_qwen3vl.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.model_executor.models.qwen3.Qwen3Attention.forward` and
 #      `vllm.model_executor.models.qwen3_moe.Qwen3MoeAttention.forward`
@@ -945,7 +885,7 @@
 #       when using mrope.
 #    Future Plan:
 #       Remove this patch when vllm-ascend supports pattern matching for this fused kernel.
-# ** 28. File: worker/patch_qwen3_dflash.py**
+# ** 26. File: worker/patch_qwen3_dflash.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.model_executor.models.qwen3_dflash.DFlashQwen3Model.precompute_and_store_context_kv`
 #    Why:
@@ -973,7 +913,7 @@
 #       Remove this patch when upstream vLLM supports MoE communication type abstraction that
 #       can be extended by hardware plugins like vllm-ascend.
 #
-# ** 29. File: platform/patch_mamba_manager.py**
+# ** 27. File: platform/patch_mamba_manager.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.core.single_type_kv_cache_manager.MambaManager`
 #    Why:

vllm_ascend/patch/worker/__init__.py

@@ -38,7 +38,6 @@
 
 
 import vllm_ascend.patch.worker.patch_process_weights_after_loading  # noqa
-import vllm_ascend.patch.worker.patch_weight_utils  # noqa
 import vllm_ascend.patch.worker.patch_distributed  # noqa
 import vllm_ascend.patch.worker.patch_minimax_m2  # noqa
 import vllm_ascend.patch.worker.patch_minimax_m2_linear_attn  # noqa
@@ -66,7 +65,6 @@
 import vllm_ascend.patch.worker.patch_cudagraph  # noqa
 import vllm_ascend.patch.worker.patch_deepseek_mtp  # noqa
 import vllm_ascend.patch.worker.patch_deepseek_v2  # noqa
-import vllm_ascend.patch.worker.patch_gqa_c8  # noqa
 
 # vLLM's use_v2_model_runner may enable the v2 runner without the
 # VLLM_USE_V2_MODEL_RUNNER env var (e.g. based on model architecture).

vllm_ascend/quantization/modelslim_config.py

@@ -531,20 +531,38 @@ def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"):
             self._add_kvcache_quant_metadata()
             logger.info("Applied hf_to_vllm_mapper to quant_description keys")
 
-    def get_cache_scale(self, name: str) -> str | None:
-        """Map checkpoint C8 KV scale/offset names to vLLM parameter names."""
-        if self.quant_description.get("kv_cache_type") != "C8":
+    def get_cache_scale_mapper(self) -> "WeightsMapper | None":
+        suffix_map = {}
+        if self.enable_c8_quant:
+            suffix_map.update(
+                {
+                    ".k_proj.kv_cache_scale": ".attn.k_cache_scale",
+                    ".k_proj.kv_cache_offset": ".attn.k_cache_offset",
+                    ".v_proj.kv_cache_scale": ".attn.v_cache_scale",
+                    ".v_proj.kv_cache_offset": ".attn.v_cache_offset",
+                }
+            )
+        if self.enable_fa_quant:
+            suffix_map.update(
+                {
+                    ".fa_q.scale": ".mla_attn.mla_attn.fa_q.scale",
+                    ".fa_k.scale": ".mla_attn.mla_attn.fa_k.scale",
+                    ".fa_v.scale": ".mla_attn.mla_attn.fa_v.scale",
+                    ".fa_q.offset": ".mla_attn.mla_attn.fa_q.offset",
+                    ".fa_k.offset": ".mla_attn.mla_attn.fa_k.offset",
+                    ".fa_v.offset": ".mla_attn.mla_attn.fa_v.offset",
+                }
+            )
+        if self.enable_indexer_quant:
+            suffix_map.update(
+                {
+                    ".indexer.q_rot": ".mla_attn.mla_attn.indexer.q_rot",
+                    ".indexer.k_rot": ".mla_attn.mla_attn.indexer.k_rot",
+                }
+            )
+        if not suffix_map:
             return None
-        _C8_SCALE_MAPPING = {
-            "k_proj.kv_cache_scale": "attn.k_cache_scale",
-            "k_proj.kv_cache_offset": "attn.k_cache_offset",
-            "v_proj.kv_cache_scale": "attn.v_cache_scale",
-            "v_proj.kv_cache_offset": "attn.v_cache_offset",
-        }
-        for src_suffix, dst_suffix in _C8_SCALE_MAPPING.items():
-            if name.endswith(src_suffix):
-                return name[: -len(src_suffix)] + dst_suffix
-        return None
+        return WeightsMapper(orig_to_new_suffix=suffix_map)
 
     def _has_quant_weight(self, prefix: str, packed_modules_mapping: Mapping[str, list[str]]) -> bool:
         proj_name = prefix.split(".")[-1]